lustre/lmv/lmv_intent.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #ifndef EXPORT_SYMTAB
  23 # define EXPORT_SYMTAB
  24 #endif
  25 #define DEBUG_SUBSYSTEM S_LMV
  26 #ifdef __KERNEL__
  27 #include <linux/slab.h>
  28 #include <linux/module.h>
  29 #include <linux/init.h>
  30 #include <linux/slab.h>
  31 #include <linux/pagemap.h>
  32 #include <asm/div64.h>
  33 #include <linux/seq_file.h>
  34 #include <linux/namei.h>
  35 #else
  36 #include <liblustre.h>
  37 #endif
  38
  39 #include <linux/obd_support.h>
  40 #include <linux/lustre_lib.h>
  41 #include <linux/lustre_net.h>
  42 #include <linux/lustre_idl.h>
  43 #include <linux/lustre_dlm.h>
  44 #include <linux/lustre_mds.h>
  45 #include <linux/obd_class.h>
  46 #include <linux/obd_ost.h>
  47 #include <linux/lprocfs_status.h>
  48 #include <linux/lustre_fsfilt.h>
  49 #include <linux/obd_lmv.h>
  50 #include <linux/lustre_lite.h>
  51 #include "lmv_internal.h"
  52
  53 static inline void lmv_drop_intent_lock(struct lookup_intent *it)
  54 {
  55         if (LUSTRE_IT(it)->it_lock_mode != 0)
  56                 ldlm_lock_decref((void *)&LUSTRE_IT(it)->it_lock_handle,
  57                                  LUSTRE_IT(it)->it_lock_mode);
  58 }
  59
  60 int lmv_intent_remote(struct obd_export *exp, void *lmm,
  61                       int lmmsize, struct lookup_intent *it,
  62                       int flags, struct ptlrpc_request **reqp,
  63                       ldlm_blocking_callback cb_blocking)
  64 {
  65         struct obd_device *obd = exp->exp_obd;
  66         struct lmv_obd *lmv = &obd->u.lmv;
  67         struct ptlrpc_request *req = NULL;
  68         struct mds_body *body = NULL;
  69         struct lustre_handle plock;
  70         struct lustre_id nid;
  71         int pmode, rc = 0;
  72         ENTRY;
  73
  74         body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
  75         LASSERT(body != NULL);
  76
  77         if (!(body->valid & OBD_MD_MDS))
  78                 RETURN(0);
  79
  80         /*
  81          * oh, MDS reports that this is remote inode case i.e. we have to ask
  82          * for real attrs on another MDS.
  83          */
  84         if (it->it_op == IT_LOOKUP || it->it_op == IT_CHDIR) {
  85                 /*
  86                  * unfortunately, we have to lie to MDC/MDS to retrieve
  87                  * attributes llite needs.
  88                  */
  89                 it->it_op = IT_GETATTR;
  90         }
  91
  92         /* we got LOOKUP lock, but we really need attrs */
  93         pmode = LUSTRE_IT(it)->it_lock_mode;
  94         if (pmode) {
  95                 memcpy(&plock, &LUSTRE_IT(it)->it_lock_handle,
  96                        sizeof(plock));
  97                 LUSTRE_IT(it)->it_lock_mode = 0;
  98                 LUSTRE_IT(it)->it_data = 0;
  99         }
 100
 101         LASSERT((body->valid & OBD_MD_FID) != 0);
 102
 103         nid = body->id1;
 104         LUSTRE_IT(it)->it_disposition &= ~DISP_ENQ_COMPLETE;
 105         rc = md_intent_lock(lmv->tgts[id_group(&nid)].ltd_exp, &nid,
 106                             NULL, 0, lmm, lmmsize, NULL, it, flags,
 107                             &req, cb_blocking);
 108
 109         /*
 110          * llite needs LOOKUP lock to track dentry revocation in order to
 111          * maintain dcache consistency. Thus drop UPDATE lock here and put
 112          * LOOKUP in request.
 113          */
 114         if (rc == 0) {
 115                 lmv_drop_intent_lock(it);
 116                 memcpy(&LUSTRE_IT(it)->it_lock_handle, &plock,
 117                        sizeof(plock));
 118                 LUSTRE_IT(it)->it_lock_mode = pmode;
 119         } else if (pmode) {
 120                 ldlm_lock_decref(&plock, pmode);
 121         }
 122
 123         ptlrpc_req_finished(*reqp);
 124         *reqp = req;
 125         RETURN(rc);
 126 }
 127
 128 int lmv_intent_open(struct obd_export *exp, struct lustre_id *pid,
 129                     const char *name, int len, void *lmm, int lmmsize,
 130                     struct lustre_id *cid, struct lookup_intent *it,
 131                     int flags, struct ptlrpc_request **reqp,
 132                     ldlm_blocking_callback cb_blocking)
 133 {
 134         struct obd_device *obd = exp->exp_obd;
 135         struct lmv_obd *lmv = &obd->u.lmv;
 136         struct mds_body *body = NULL;
 137         struct lustre_id rpid = *pid;
 138         int rc, mds, loop = 0;
 139         struct lmv_obj *obj;
 140         struct mea *mea;
 141         ENTRY;
 142
 143         /* IT_OPEN is intended to open (and create, possible) an object. Parent
 144          * (pid) may be splitted dir */
 145
 146 repeat:
 147         LASSERT(++loop <= 2);
 148         mds = id_group(&rpid);
 149         obj = lmv_grab_obj(obd, &rpid);
 150         if (obj) {
 151                 /* directory is already splitted, so we have to forward
 152                  * request to the right MDS */
 153                 mds = raw_name2idx(obj->hashtype, obj->objcount,
 154                                    (char *)name, len);
 155
 156                 CDEBUG(D_OTHER, "forward to MDS #%u ("DLID4")\n",
 157                        mds, OLID4(&rpid));
 158                 rpid = obj->objs[mds].id;
 159                 lmv_put_obj(obj);
 160         }
 161
 162         rc = md_intent_lock(lmv->tgts[id_group(&rpid)].ltd_exp, &rpid, name,
 163                             len, lmm, lmmsize, cid, it, flags, reqp, cb_blocking);
 164         if (rc == -ERESTART) {
 165                 /* directory got splitted. time to update local object and
 166                  * repeat the request with proper MDS */
 167                 LASSERT(id_equal_fid(pid, &rpid));
 168                 rc = lmv_get_mea_and_update_object(exp, &rpid);
 169                 if (rc == 0) {
 170                         ptlrpc_req_finished(*reqp);
 171                         goto repeat;
 172                 }
 173         }
 174         if (rc != 0)
 175                 RETURN(rc);
 176
 177         /* okay, MDS has returned success. Probably name has been resolved in
 178          * remote inode */
 179         rc = lmv_intent_remote(exp, lmm, lmmsize, it, flags, reqp, cb_blocking);
 180         if (rc != 0) {
 181                 LASSERT(rc < 0);
 182
 183                 /*
 184                  * this is possible, that some userspace application will try to
 185                  * open file as directory and we will have -ENOTDIR here. As
 186                  * this is "usual" situation, we should not print error here,
 187                  * only debug info.
 188                  */
 189                 CDEBUG(D_OTHER, "can't handle remote %s: dir "DLID4"("DLID4"):"
 190                        "%*s: %d\n", LL_IT2STR(it), OLID4(pid), OLID4(&rpid),
 191                        len, name, rc);
 192                 RETURN(rc);
 193         }
 194
 195         /*
 196          * nothing is found, do not access body->id1 as it is zero and thus
 197          * pointless.
 198          */
 199         if ((LUSTRE_IT(it)->it_disposition & DISP_LOOKUP_NEG) &&
 200             !(LUSTRE_IT(it)->it_disposition & DISP_OPEN_CREATE) &&
 201             !(LUSTRE_IT(it)->it_disposition & DISP_OPEN_OPEN))
 202                 RETURN(0);
 203
 204         /* caller may use attrs MDS returns on IT_OPEN lock request so, we have
 205          * to update them for splitted dir */
 206         body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 207         LASSERT(body != NULL);
 208
 209         /* could not find object, FID is not present in response. */
 210         if (!(body->valid & OBD_MD_FID))
 211                 RETURN(0);
 212
 213         cid = &body->id1;
 214         obj = lmv_grab_obj(obd, cid);
 215         if (!obj && (mea = lmv_splitted_dir_body(*reqp, 1))) {
 216                 /* wow! this is splitted dir, we'd like to handle it */
 217                 obj = lmv_create_obj(exp, &body->id1, mea);
 218                 if (IS_ERR(obj))
 219                         RETURN(PTR_ERR(obj));
 220         }
 221
 222         if (obj) {
 223                 /* this is splitted dir and we'd want to get attrs */
 224                 CDEBUG(D_OTHER, "attrs from slaves for "DLID4"\n",
 225                        OLID4(cid));
 226
 227                 rc = lmv_revalidate_slaves(exp, reqp, cid, it, 1,
 228                                            cb_blocking);
 229         } else if (S_ISDIR(body->mode)) {
 230                 CDEBUG(D_OTHER, "object "DLID4" has not lmv obj?\n",
 231                        OLID4(cid));
 232         }
 233
 234         if (obj)
 235                 lmv_put_obj(obj);
 236
 237         RETURN(rc);
 238 }
 239
 240 int lmv_intent_getattr(struct obd_export *exp, struct lustre_id *pid,
 241                        const char *name, int len, void *lmm, int lmmsize,
 242                        struct lustre_id *cid, struct lookup_intent *it,
 243                        int flags, struct ptlrpc_request **reqp,
 244                        ldlm_blocking_callback cb_blocking)
 245 {
 246         struct obd_device *obd = exp->exp_obd;
 247         struct lmv_obd *lmv = &obd->u.lmv;
 248         struct mds_body *body = NULL;
 249         struct lustre_id rpid = *pid;
 250         struct lmv_obj *obj = NULL, *obj2 = NULL;
 251         struct mea *mea;
 252         int rc = 0, mds;
 253         ENTRY;
 254
 255         if (cid) {
 256                 /* caller wants to revalidate attrs of obj we have to revalidate
 257                  * slaves if requested object is splitted directory */
 258                 CDEBUG(D_OTHER, "revalidate attrs for "DLID4"\n", OLID4(cid));
 259                 mds = id_group(cid);
 260 #if 0
 261                 obj = lmv_grab_obj(obd, cid);
 262                 if (obj) {
 263                         /* in fact, we need not this with current intent_lock(),
 264                          * but it may change some day */
 265                         if (!id_equal_fid(pid, cid)){
 266                                 rpid = obj->objs[mds].id;
 267                                 mds = id_group(&rpid);
 268                         }
 269                         lmv_put_obj(obj);
 270                 }
 271 #endif
 272         } else {
 273                 CDEBUG(D_OTHER, "INTENT getattr for %*s on "DLID4"\n",
 274                        len, name, OLID4(pid));
 275                 mds = id_group(pid);
 276                 obj = lmv_grab_obj(obd, pid);
 277                 if (obj && len) {
 278                         /* directory is already splitted. calculate mds */
 279                         mds = raw_name2idx(obj->hashtype, obj->objcount,
 280                                            (char *)name, len);
 281                         rpid = obj->objs[mds].id;
 282                         mds = id_group(&rpid);
 283                         lmv_put_obj(obj);
 284
 285                         CDEBUG(D_OTHER, "forward to MDS #%u (slave "DLID4")\n",
 286                                mds, OLID4(&rpid));
 287                 }
 288         }
 289
 290         /* the same about fid returning. */
 291         rc = md_intent_lock(lmv->tgts[mds].ltd_exp, &rpid, name, len, lmm,
 292                             lmmsize, cid, it, flags, reqp, cb_blocking);
 293         if (rc < 0)
 294                 RETURN(rc);
 295
 296         if (obj && rc > 0) {
 297                 /* this is splitted dir. In order to optimize things a
 298                  * bit, we consider obj valid updating missing parts.
 299
 300                  * FIXME: do we need to return any lock here? It would
 301                  * be fine if we don't. this means that nobody should
 302                  * use UPDATE lock to notify about object * removal */
 303                 CDEBUG(D_OTHER,
 304                        "revalidate slaves for "DLID4", rc %d\n",
 305                        OLID4(cid), rc);
 306
 307                 LASSERT(cid != 0);
 308                 rc = lmv_revalidate_slaves(exp, reqp, cid, it, rc,
 309                                            cb_blocking);
 310                 RETURN(rc);
 311         }
 312
 313         if (*reqp == NULL)
 314                 RETURN(rc);
 315
 316         /* okay, MDS has returned success. probably name has been
 317          * resolved in remote inode */
 318         rc = lmv_intent_remote(exp, lmm, lmmsize, it, flags,
 319                                reqp, cb_blocking);
 320         if (rc < 0)
 321                 RETURN(rc);
 322
 323         /*
 324          * nothing is found, do not access body->id1 as it is zero and thus
 325          * pointless.
 326          */
 327         if (LUSTRE_IT(it)->it_disposition & DISP_LOOKUP_NEG)
 328                 RETURN(0);
 329
 330         body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 331         LASSERT(body != NULL);
 332
 333         /* could not find object, FID is not present in response. */
 334         if (!(body->valid & OBD_MD_FID))
 335                 RETURN(0);
 336
 337         cid = &body->id1;
 338         obj2 = lmv_grab_obj(obd, cid);
 339
 340         if (!obj2 && (mea = lmv_splitted_dir_body(*reqp, 1))) {
 341                 /* wow! this is splitted dir, we'd like to handle it. */
 342                 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 343                 LASSERT(body != NULL);
 344
 345                 obj2 = lmv_create_obj(exp, &body->id1, mea);
 346                 if (IS_ERR(obj2))
 347                         RETURN(PTR_ERR(obj2));
 348         }
 349
 350         if (obj2) {
 351                 /* this is splitted dir and we'd want to get attrs */
 352                 CDEBUG(D_OTHER, "attrs from slaves for "DLID4", rc %d\n",
 353                        OLID4(cid), rc);
 354
 355                 rc = lmv_revalidate_slaves(exp, reqp, cid, it, 1,
 356                                            cb_blocking);
 357                 lmv_put_obj(obj2);
 358         }
 359         RETURN(rc);
 360 }
 361
 362 void lmv_update_body_from_obj(struct mds_body *body, struct lmv_inode *obj)
 363 {
 364         /* update size */
 365         body->size += obj->size;
 366 }
 367
 368 int lmv_lookup_slaves(struct obd_export *exp, struct ptlrpc_request **reqp)
 369 {
 370         struct obd_device *obd = exp->exp_obd;
 371         struct lmv_obd *lmv = &obd->u.lmv;
 372         struct mds_body *body = NULL;
 373         struct lustre_handle *lockh;
 374         struct ldlm_lock *lock;
 375         struct mds_body *body2;
 376         struct lmv_obj *obj;
 377         int i, rc = 0;
 378         ENTRY;
 379
 380         LASSERT(reqp);
 381         LASSERT(*reqp);
 382
 383         /* master is locked. we'd like to take locks on slaves and update
 384          * attributes to be returned from the slaves it's important that lookup
 385          * is called in two cases:
 386
 387          *  - for first time (dcache has no such a resolving yet).
 388          *  - ->d_revalidate() returned false.
 389
 390          * last case possible only if all the objs (master and all slaves aren't
 391          * valid */
 392
 393         body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 394         LASSERT(body != NULL);
 395         LASSERT((body->valid & OBD_MD_FID) != 0);
 396
 397         obj = lmv_grab_obj(obd, &body->id1);
 398         LASSERT(obj != NULL);
 399
 400         CDEBUG(D_OTHER, "lookup slaves for "DLID4"\n",
 401                OLID4(&body->id1));
 402
 403         lmv_lock_obj(obj);
 404
 405         for (i = 0; i < obj->objcount; i++) {
 406                 struct lustre_id id = obj->objs[i].id;
 407                 struct ptlrpc_request *req = NULL;
 408                 struct lookup_intent it;
 409
 410                 if (id_equal_fid(&id, &obj->id))
 411                         /* skip master obj */
 412                         continue;
 413
 414                 CDEBUG(D_OTHER, "lookup slave "DLID4"\n", OLID4(&id));
 415
 416                 /* is obj valid? */
 417                 memset(&it, 0, sizeof(it));
 418                 it.it_op = IT_GETATTR;
 419                 OBD_ALLOC(it.d.fs_data, sizeof(struct lustre_intent_data));
 420
 421                 rc = md_intent_lock(lmv->tgts[id_group(&id)].ltd_exp, &id,
 422                                     NULL, 0, NULL, 0, &id, &it, 0, &req,
 423                                     lmv_dirobj_blocking_ast);
 424
 425                 lockh = (struct lustre_handle *)&LUSTRE_IT(&it)->it_lock_handle;
 426                 if (rc > 0 && req == NULL) {
 427                         /* nice, this slave is valid */
 428                         LASSERT(req == NULL);
 429                         CDEBUG(D_OTHER, "cached\n");
 430                         goto release_lock;
 431                 }
 432
 433                 if (rc < 0) {
 434                         OBD_FREE(it.d.fs_data, sizeof(struct lustre_intent_data));
 435                         /* error during lookup */
 436                         GOTO(cleanup, rc);
 437                 }
 438                 lock = ldlm_handle2lock(lockh);
 439                 LASSERT(lock);
 440
 441                 lock->l_ast_data = lmv_get_obj(obj);
 442
 443                 body2 = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body2));
 444                 LASSERT(body2);
 445
 446                 obj->objs[i].size = body2->size;
 447
 448                 CDEBUG(D_OTHER, "fresh: %lu\n",
 449                        (unsigned long)obj->objs[i].size);
 450
 451                 LDLM_LOCK_PUT(lock);
 452
 453                 if (req)
 454                         ptlrpc_req_finished(req);
 455 release_lock:
 456                 lmv_update_body_from_obj(body, obj->objs + i);
 457
 458                 if (LUSTRE_IT(&it)->it_lock_mode)
 459                         ldlm_lock_decref(lockh, LUSTRE_IT(&it)->it_lock_mode);
 460                 OBD_FREE(it.d.fs_data, sizeof(struct lustre_intent_data));
 461         }
 462
 463         EXIT;
 464 cleanup:
 465         lmv_unlock_obj(obj);
 466         lmv_put_obj(obj);
 467         return rc;
 468 }
 469
 470 int lmv_intent_lookup(struct obd_export *exp, struct lustre_id *pid,
 471                       const char *name, int len, void *lmm, int lmmsize,
 472                       struct lustre_id *cid, struct lookup_intent *it,
 473                       int flags, struct ptlrpc_request **reqp,
 474                       ldlm_blocking_callback cb_blocking)
 475 {
 476         struct obd_device *obd = exp->exp_obd;
 477         struct lmv_obd *lmv = &obd->u.lmv;
 478         struct mds_body *body = NULL;
 479         struct lustre_id rpid = *pid;
 480         struct lmv_obj *obj;
 481         struct mea *mea;
 482         int rc, mds, loop = 0;
 483         ENTRY;
 484
 485         /*
 486          * IT_LOOKUP is intended to produce name -> id resolving (let's call
 487          * this lookup below) or to confirm requested resolving is still valid
 488          * (let's call this revalidation) cid != NULL specifies revalidation.
 489          */
 490         if (cid) {
 491                 /*
 492                  * this is revalidation: we have to check is LOOKUP lock still
 493                  * valid for given id. Very important part is that we have to
 494                  * choose right mds because namespace is per mds.
 495                  */
 496                 rpid = *pid;
 497                 obj = lmv_grab_obj(obd, pid);
 498                 if (obj) {
 499                         mds = raw_name2idx(obj->hashtype, obj->objcount,
 500                                            (char *)name, len);
 501                         rpid = obj->objs[mds].id;
 502                         lmv_put_obj(obj);
 503                 }
 504                 mds = id_group(&rpid);
 505
 506                 CDEBUG(D_OTHER, "revalidate lookup for "DLID4" to %d MDS\n",
 507                        OLID4(cid), mds);
 508
 509         } else {
 510                 mds = id_group(pid);
 511 repeat:
 512                 LASSERT(++loop <= 2);
 513
 514                 /* this is lookup. during lookup we have to update all the
 515                  * attributes, because returned values will be put in struct
 516                  * inode */
 517
 518                 obj = lmv_grab_obj(obd, pid);
 519                 if (obj) {
 520                         if (len) {
 521                                 /* directory is already splitted. calculate mds */
 522                                 mds = raw_name2idx(obj->hashtype, obj->objcount,
 523                                                    (char *)name, len);
 524                                 rpid = obj->objs[mds].id;
 525                                 mds = id_group(&rpid);
 526                         }
 527                         lmv_put_obj(obj);
 528                 }
 529         }
 530         rc = md_intent_lock(lmv->tgts[mds].ltd_exp, &rpid, name,
 531                             len, lmm, lmmsize, cid, it, flags,
 532                             reqp, cb_blocking);
 533         if (rc > 0) {
 534                 LASSERT(cid != 0);
 535                 RETURN(rc);
 536         }
 537         if (rc > 0) {
 538                 /* very interesting. it seems object is still valid but for some
 539                  * reason llite calls lookup, not revalidate */
 540                 CDEBUG(D_OTHER, "lookup for "DLID4" and data should be uptodate\n",
 541                       OLID4(&rpid));
 542                 LASSERT(*reqp == NULL);
 543                 RETURN(rc);
 544         }
 545
 546         if (rc == 0 && *reqp == NULL) {
 547                 /* once again, we're asked for lookup, not revalidate */
 548                 CDEBUG(D_OTHER, "lookup for "DLID4" and data should be uptodate\n",
 549                       OLID4(&rpid));
 550                 RETURN(rc);
 551         }
 552
 553         if (rc == -ERESTART) {
 554                 /* directory got splitted since last update. this shouldn't be
 555                  * becasue splitting causes lock revocation, so revalidate had
 556                  * to fail and lookup on dir had to return mea */
 557                 CWARN("we haven't knew about directory splitting!\n");
 558                 LASSERT(obj == NULL);
 559
 560                 obj = lmv_create_obj(exp, &rpid, NULL);
 561                 if (IS_ERR(obj))
 562                         RETURN(PTR_ERR(obj));
 563                 lmv_put_obj(obj);
 564                 goto repeat;
 565         }
 566
 567         if (rc < 0)
 568                 RETURN(rc);
 569
 570         /* okay, MDS has returned success. Probably name has been resolved in
 571          * remote inode. */
 572         rc = lmv_intent_remote(exp, lmm, lmmsize, it, flags, reqp, cb_blocking);
 573
 574         if (rc == 0 && (mea = lmv_splitted_dir_body(*reqp, 1))) {
 575                 /* wow! this is splitted dir, we'd like to handle it */
 576                 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 577                 LASSERT(body != NULL);
 578                 LASSERT((body->valid & OBD_MD_FID) != 0);
 579
 580                 obj = lmv_grab_obj(obd, &body->id1);
 581                 if (!obj) {
 582                         obj = lmv_create_obj(exp, &body->id1, mea);
 583                         if (IS_ERR(obj))
 584                                 RETURN(PTR_ERR(obj));
 585                 }
 586                 lmv_put_obj(obj);
 587         }
 588
 589         RETURN(rc);
 590 }
 591
 592 int lmv_intent_lock(struct obd_export *exp, struct lustre_id *pid,
 593                     const char *name, int len, void *lmm, int lmmsize,
 594                     struct lustre_id *cid, struct lookup_intent *it,
 595                     int flags, struct ptlrpc_request **reqp,
 596                     ldlm_blocking_callback cb_blocking)
 597 {
 598         struct obd_device *obd = exp->exp_obd;
 599         int rc = 0;
 600         ENTRY;
 601
 602         LASSERT(it);
 603         LASSERT(pid);
 604
 605         CDEBUG(D_OTHER, "INTENT LOCK '%s' for '%*s' on %lu/%lu -> %lu\n",
 606                LL_IT2STR(it), len, name, (unsigned long)id_ino(pid),
 607                (unsigned long)id_gen(pid), (unsigned long)id_group(pid));
 608
 609         rc = lmv_check_connect(obd);
 610         if (rc)
 611                 RETURN(rc);
 612
 613         if (it->it_op == IT_LOOKUP)
 614                 rc = lmv_intent_lookup(exp, pid, name, len, lmm,
 615                                        lmmsize, cid, it, flags, reqp,
 616                                        cb_blocking);
 617         else if (it->it_op & IT_OPEN)
 618                 rc = lmv_intent_open(exp, pid, name, len, lmm,
 619                                      lmmsize, cid, it, flags, reqp,
 620                                      cb_blocking);
 621         else if (it->it_op == IT_GETATTR || it->it_op == IT_CHDIR)
 622                 rc = lmv_intent_getattr(exp, pid, name, len, lmm,
 623                                         lmmsize, cid, it, flags, reqp,
 624                                         cb_blocking);
 625         else
 626                 LBUG();
 627         RETURN(rc);
 628 }
 629
 630 int lmv_revalidate_slaves(struct obd_export *exp, struct ptlrpc_request **reqp,
 631                           struct lustre_id *mid, struct lookup_intent *oit,
 632                           int master_valid, ldlm_blocking_callback cb_blocking)
 633 {
 634         struct obd_device *obd = exp->exp_obd;
 635         struct ptlrpc_request *mreq = *reqp;
 636         struct lmv_obd *lmv = &obd->u.lmv;
 637         struct lustre_handle master_lockh;
 638         struct ldlm_lock *lock;
 639         unsigned long size = 0;
 640         struct mds_body *body;
 641         struct lmv_obj *obj;
 642         int master_lock_mode;
 643         int i, rc = 0;
 644         ENTRY;
 645
 646         /* we have to loop over the subobjects, check validity and update them
 647          * from MDSs if needed. it's very useful that we need not to update all
 648          * the fields. say, common fields (that are equal on all the subojects
 649          * need not to be update, another fields (i_size, for example) are
 650          * cached all the time */
 651         obj = lmv_grab_obj(obd, mid);
 652         LASSERT(obj != NULL);
 653
 654         master_lock_mode = 0;
 655
 656         lmv_lock_obj(obj);
 657
 658         for (i = 0; i < obj->objcount; i++) {
 659                 struct lustre_id id = obj->objs[i].id;
 660                 struct lustre_handle *lockh = NULL;
 661                 struct ptlrpc_request *req = NULL;
 662                 ldlm_blocking_callback cb;
 663                 struct lookup_intent it;
 664                 int master = 0;
 665
 666                 CDEBUG(D_OTHER, "revalidate subobj "DLID4"\n",
 667                        OLID4(&id));
 668
 669                 memset(&it, 0, sizeof(it));
 670                 it.it_op = IT_GETATTR;
 671
 672                 cb = lmv_dirobj_blocking_ast;
 673
 674                 OBD_ALLOC(it.d.fs_data, sizeof(struct lustre_intent_data));
 675                 if (id_equal_fid(&id, &obj->id)) {
 676                         if (master_valid) {
 677                                 /* lmv_intent_getattr() already checked
 678                                  * validness and took the lock */
 679                                 if (mreq) {
 680                                         /* it even got the reply refresh attrs
 681                                          * from that reply */
 682                                         body = lustre_msg_buf(mreq->rq_repmsg,
 683                                                               1, sizeof(*body));
 684                                         LASSERT(body != NULL);
 685                                         goto update;
 686                                 }
 687                                 /* take already cached attrs into account */
 688                                 CDEBUG(D_OTHER,
 689                                        "master is locked and cached\n");
 690                                 goto release_lock;
 691                         }
 692                         master = 1;
 693                         cb = cb_blocking;
 694                 }
 695
 696
 697                 /* is obj valid? */
 698                 rc = md_intent_lock(lmv->tgts[id_group(&id)].ltd_exp,
 699                                     &id, NULL, 0, NULL, 0, &id, &it, 0,
 700                                     &req, cb);
 701                 lockh = (struct lustre_handle *) &LUSTRE_IT(&it)->it_lock_handle;
 702                 if (rc > 0 && req == NULL) {
 703                         /* nice, this slave is valid */
 704                         LASSERT(req == NULL);
 705                         CDEBUG(D_OTHER, "cached\n");
 706                         goto release_lock;
 707                 }
 708
 709                 if (rc < 0) {
 710                         OBD_FREE(it.d.fs_data, sizeof(struct lustre_intent_data));
 711                         /* error during revalidation */
 712                         GOTO(cleanup, rc);
 713                 }
 714                 if (master) {
 715                         LASSERT(master_valid == 0);
 716                         /* save lock on master to be returned to the caller */
 717                         CDEBUG(D_OTHER, "no lock on master yet\n");
 718                         memcpy(&master_lockh, lockh, sizeof(master_lockh));
 719                         master_lock_mode = LUSTRE_IT(&it)->it_lock_mode;
 720                         LUSTRE_IT(&it)->it_lock_mode = 0;
 721                 } else {
 722                         /* this is slave. we want to control it */
 723                         lock = ldlm_handle2lock(lockh);
 724                         LASSERT(lock);
 725                         lock->l_ast_data = lmv_get_obj(obj);
 726                         LDLM_LOCK_PUT(lock);
 727                 }
 728
 729                 if (*reqp == NULL) {
 730                         /* this is first reply, we'll use it to return updated
 731                          * data back to the caller */
 732                         LASSERT(req);
 733                         ptlrpc_request_addref(req);
 734                         *reqp = req;
 735
 736                 }
 737
 738                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
 739                 LASSERT(body);
 740
 741 update:
 742                 obj->objs[i].size = body->size;
 743
 744                 CDEBUG(D_OTHER, "fresh: %lu\n",
 745                        (unsigned long)obj->objs[i].size);
 746
 747                 if (req)
 748                         ptlrpc_req_finished(req);
 749 release_lock:
 750                 size += obj->objs[i].size;
 751
 752                 if (LUSTRE_IT(&it)->it_lock_mode)
 753                         ldlm_lock_decref(lockh, LUSTRE_IT(&it)->it_lock_mode);
 754                 OBD_FREE(it.d.fs_data, sizeof(struct lustre_intent_data));
 755         }
 756
 757         if (*reqp) {
 758                 /* some attrs got refreshed, we have reply and it's time to put
 759                  * fresh attrs to it */
 760                 CDEBUG(D_OTHER, "return refreshed attrs: size = %lu\n",
 761                        (unsigned long)size);
 762
 763                 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 764                 LASSERT(body);
 765
 766                 /* FIXME: what about other attributes? */
 767                 body->size = size;
 768
 769                 if (mreq == NULL) {
 770                         /* very important to maintain id_group(lli->lli_id) the
 771                          * same because of revalidation. mreq == NULL means that
 772                          * caller has no reply and the only attr we can return
 773                          * is size */
 774                         body->valid = OBD_MD_FLSIZE;
 775 //                        body->mds = id_group(&obj->id);
 776                 }
 777                 if (master_valid == 0) {
 778                         memcpy(&LUSTRE_IT(oit)->it_lock_handle,
 779                                &master_lockh, sizeof(master_lockh));
 780                         LUSTRE_IT(oit)->it_lock_mode = master_lock_mode;
 781                 }
 782                 rc = 0;
 783         } else {
 784                 /* it seems all the attrs are fresh and we did no request */
 785                 CDEBUG(D_OTHER, "all the attrs were fresh\n");
 786                 if (master_valid == 0)
 787                         LUSTRE_IT(oit)->it_lock_mode = master_lock_mode;
 788                 rc = 1;
 789         }
 790
 791         EXIT;
 792 cleanup:
 793         lmv_unlock_obj(obj);
 794         lmv_put_obj(obj);
 795         return rc;
 796 }