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