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