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,
  60                             void *lmm, int lmmsize,
  61                             struct lookup_intent *it, int flags,
  62                             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 mds_body *body = NULL;
  68         int rc = 0;
  69         ENTRY;
  70
  71         body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
  72         LASSERT(body != NULL);
  73
  74         if (body->valid & OBD_MD_MDS) {
  75                 /* oh, MDS reports that this is remote inode case
  76                  * i.e. we have to ask for real attrs on another MDS */
  77                 struct ptlrpc_request *req = NULL;
  78                 struct ll_fid nfid;
  79                 struct lustre_handle plock;
  80                 int pmode;
  81
  82                 if (it->it_op == IT_LOOKUP) {
  83                         /* unfortunately, we have to lie to MDC/MDS to
  84                          * retrieve attributes llite needs */
  85                         it->it_op = IT_GETATTR;
  86                 }
  87
  88                 /* we got LOOKUP lock, but we really need attrs */
  89                 pmode = it->d.lustre.it_lock_mode;
  90                 if (pmode) {
  91                         memcpy(&plock, &it->d.lustre.it_lock_handle,
  92                                         sizeof(plock));
  93                         it->d.lustre.it_lock_mode = 0;
  94                 }
  95
  96                 nfid = body->fid1;
  97                 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
  98                 rc = md_intent_lock(lmv->tgts[nfid.mds].ltd_exp, &nfid,
  99                                     NULL, 0, lmm, lmmsize, NULL, it, flags,
 100                                     &req, cb_blocking);
 101
 102                 /* llite needs LOOKUP lock to track dentry revocation in
 103                  * order to maintain dcache consistency. thus drop UPDATE
 104                  * lock here and put LOOKUP in request */
 105                 if (rc == 0) {
 106                         lmv_drop_intent_lock(it);
 107                         memcpy(&it->d.lustre.it_lock_handle, &plock,
 108                                         sizeof(plock));
 109                         it->d.lustre.it_lock_mode = pmode;
 110
 111                 } else if (pmode)
 112                         ldlm_lock_decref(&plock, pmode);
 113
 114                 ptlrpc_req_finished(*reqp);
 115                 *reqp = req;
 116         }
 117         RETURN(rc);
 118 }
 119
 120 int lmv_intent_open(struct obd_export *exp,
 121                     struct ll_fid *pfid, const char *name, int len,
 122                     void *lmm, int lmmsize, struct ll_fid *cfid,
 123                     struct lookup_intent *it, int flags,
 124                     struct ptlrpc_request **reqp,
 125                     ldlm_blocking_callback cb_blocking)
 126 {
 127         struct obd_device *obd = exp->exp_obd;
 128         struct lmv_obd *lmv = &obd->u.lmv;
 129         struct mds_body *body = NULL;
 130         struct ll_fid rpfid = *pfid;
 131         struct lmv_obj *obj;
 132         struct mea *mea;
 133         int rc, mds, loop = 0;
 134         ENTRY;
 135
 136         /* IT_OPEN is intended to open (and create, possible) an object. Parent
 137          * (pfid) may be splitted dir */
 138
 139 repeat:
 140         LASSERT(++loop <= 2);
 141         mds = rpfid.mds;
 142         obj = lmv_grab_obj(obd, &rpfid);
 143         if (obj) {
 144                 /* directory is already splitted, so we have to forward
 145                  * request to the right MDS */
 146                 mds = raw_name2idx(obj->hashtype, obj->objcount, (char *)name, len);
 147                 CDEBUG(D_OTHER, "forward to MDS #%u (%lu/%lu/%lu)\n", mds,
 148                        (unsigned long) rpfid.mds, (unsigned long) rpfid.id,
 149                        (unsigned long) rpfid.generation);
 150                 rpfid = obj->objs[mds].fid;
 151                 lmv_put_obj(obj);
 152         }
 153
 154         rc = md_intent_lock(lmv->tgts[rpfid.mds].ltd_exp, &rpfid, name,
 155                             len, lmm, lmmsize, cfid, it, flags, reqp,
 156                             cb_blocking);
 157         if (rc == -ERESTART) {
 158                 /* directory got splitted. time to update local object
 159                  * and repeat the request with proper MDS */
 160                 LASSERT(fid_equal(pfid, &rpfid));
 161                 rc = lmv_get_mea_and_update_object(exp, &rpfid);
 162                 if (rc == 0) {
 163                         ptlrpc_req_finished(*reqp);
 164                         goto repeat;
 165                 }
 166         }
 167         if (rc != 0)
 168                 RETURN(rc);
 169
 170         /* okay, MDS has returned success. Probably name has been resolved in
 171          * remote inode */
 172         rc = lmv_handle_remote_inode(exp, lmm, lmmsize, it,
 173                                      flags, reqp, cb_blocking);
 174         if (rc != 0) {
 175                 LASSERT(rc < 0);
 176                 CERROR("can't handle remote %s: dir %lu/%lu/%lu(%lu/%lu/%lu):"
 177                        "%*s: %d\n", LL_IT2STR(it),
 178                        (unsigned long) pfid->mds,
 179                        (unsigned long) pfid->id,
 180                        (unsigned long) pfid->generation,
 181                        (unsigned long) rpfid.mds,
 182                        (unsigned long) rpfid.id,
 183                        (unsigned long) rpfid.generation,
 184                        len, name, rc);
 185                 RETURN(rc);
 186         }
 187
 188         /* caller may use attrs MDS returns on IT_OPEN lock request so, we have
 189          * to update them for splitted dir */
 190         body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 191         LASSERT(body != NULL);
 192
 193         cfid = &body->fid1;
 194         obj = lmv_grab_obj(obd, cfid);
 195         if (!obj && (mea = body_of_splitted_dir(*reqp, 1))) {
 196                 /* wow! this is splitted dir, we'd like to handle it */
 197                 obj = lmv_create_obj(exp, &body->fid1, mea);
 198                 if (IS_ERR(obj))
 199                         RETURN(PTR_ERR(obj));
 200         }
 201
 202         if (obj) {
 203                 /* this is splitted dir and we'd want to get attrs */
 204                 CDEBUG(D_OTHER, "attrs from slaves for %lu/%lu/%lu\n",
 205                        (unsigned long)cfid->mds, (unsigned long)cfid->id,
 206                        (unsigned long)cfid->generation);
 207                 rc = lmv_revalidate_slaves(exp, reqp, cfid,
 208                                            it, 1, cb_blocking);
 209         } else if (S_ISDIR(body->mode)) {
 210                 /*CWARN("hmmm, %lu/%lu/%lu has not lmv obj?!\n",
 211                                 (unsigned long) cfid->mds,
 212                                 (unsigned long) cfid->id,
 213                                 (unsigned long) cfid->generation);*/
 214         }
 215
 216         if (obj)
 217                 lmv_put_obj(obj);
 218
 219         RETURN(rc);
 220 }
 221
 222 int lmv_intent_getattr(struct obd_export *exp,
 223                        struct ll_fid *pfid, const char *name, int len,
 224                        void *lmm, int lmmsize, struct ll_fid *cfid,
 225                        struct lookup_intent *it, int flags,
 226                        struct ptlrpc_request **reqp,
 227                        ldlm_blocking_callback cb_blocking)
 228 {
 229         struct obd_device *obd = exp->exp_obd;
 230         struct lmv_obd *lmv = &obd->u.lmv;
 231         struct mds_body *body = NULL;
 232         struct ll_fid rpfid = *pfid;
 233         struct lmv_obj *obj, *obj2;
 234         struct mea *mea;
 235         int rc = 0, mds;
 236         ENTRY;
 237
 238         if (cfid) {
 239                 /* caller wants to revalidate attrs of obj we have to revalidate
 240                  * slaves if requested object is splitted directory */
 241                 CDEBUG(D_OTHER, "revalidate attrs for %lu/%lu/%lu\n",
 242                        (unsigned long)cfid->mds, (unsigned long)cfid->id,
 243                        (unsigned long)cfid->generation);
 244                 mds = cfid->mds;
 245                 obj = lmv_grab_obj(obd, cfid);
 246                 if (obj) {
 247                         /* in fact, we need not this with current intent_lock(),
 248                          * but it may change some day */
 249                         rpfid = obj->objs[mds].fid;
 250                         mds = rpfid.mds;
 251                         lmv_put_obj(obj);
 252                 }
 253                 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, &rpfid, name,
 254                                     len, lmm, lmmsize, cfid, it, flags, reqp,
 255                                     cb_blocking);
 256                 if (obj && rc >= 0) {
 257                         /* this is splitted dir. In order to optimize things a
 258                          * bit, we consider obj valid updating missing parts.
 259
 260                          * FIXME: do we need to return any lock here? It would
 261                          * be fine if we don't. this means that nobody should
 262                          * use UPDATE lock to notify about object * removal */
 263                         CDEBUG(D_OTHER,
 264                                "revalidate slaves for %lu/%lu/%lu, rc %d\n",
 265                                (unsigned long)cfid->mds, (unsigned long)cfid->id,
 266                                (unsigned long)cfid->generation, rc);
 267
 268                         rc = lmv_revalidate_slaves(exp, reqp, cfid, it, rc,
 269                                                    cb_blocking);
 270                 }
 271
 272                 RETURN(rc);
 273         }
 274
 275         CDEBUG(D_OTHER, "INTENT getattr for %*s on %lu/%lu/%lu\n",
 276                len, name, (unsigned long)pfid->mds, (unsigned long)pfid->id,
 277                (unsigned long)pfid->generation);
 278
 279         mds = pfid->mds;
 280         obj = lmv_grab_obj(obd, pfid);
 281         if (obj && len) {
 282                 /* directory is already splitted. calculate mds */
 283                 mds = raw_name2idx(obj->hashtype, obj->objcount, (char *) name, len);
 284                 rpfid = obj->objs[mds].fid;
 285                 mds = rpfid.mds;
 286                 lmv_put_obj(obj);
 287
 288                 CDEBUG(D_OTHER, "forward to MDS #%u (slave %lu/%lu/%lu)\n",
 289                        mds, (unsigned long)rpfid.mds, (unsigned long)rpfid.id,
 290                        (unsigned long)rpfid.generation);
 291         }
 292
 293         rc = md_intent_lock(lmv->tgts[mds].ltd_exp, &rpfid, name,
 294                             len, lmm, lmmsize, NULL, it, flags, reqp,
 295                             cb_blocking);
 296
 297         if (rc < 0)
 298                 RETURN(rc);
 299
 300         LASSERT(rc == 0);
 301
 302         /* okay, MDS has returned success. probably name has been
 303          * resolved in remote inode */
 304         rc = lmv_handle_remote_inode(exp, lmm, lmmsize, it,
 305                                      flags, reqp, cb_blocking);
 306         if (rc < 0)
 307                 RETURN(rc);
 308
 309         body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 310         LASSERT(body != NULL);
 311
 312         cfid = &body->fid1;
 313         obj2 = lmv_grab_obj(obd, cfid);
 314
 315         if (!obj2 && (mea = body_of_splitted_dir(*reqp, 1))) {
 316                 /* wow! this is splitted dir, we'd like to handle it. */
 317                 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 318                 LASSERT(body != NULL);
 319
 320                 obj2 = lmv_create_obj(exp, &body->fid1, mea);
 321                 if (IS_ERR(obj2))
 322                         RETURN(PTR_ERR(obj2));
 323         }
 324
 325         if (obj2) {
 326                 /* this is splitted dir and we'd want to get attrs */
 327                 CDEBUG(D_OTHER, "attrs from slaves for %lu/%lu/%lu, rc %d\n",
 328                        (unsigned long)cfid->mds, (unsigned long)cfid->id,
 329                        (unsigned long)cfid->generation, rc);
 330
 331                 rc = lmv_revalidate_slaves(exp, reqp, cfid, it, 1, cb_blocking);
 332                 lmv_put_obj(obj2);
 333         }
 334         RETURN(rc);
 335 }
 336
 337 void lmv_update_body_from_obj(struct mds_body *body, struct lmv_inode *obj)
 338 {
 339         /* update size */
 340         body->size += obj->size;
 341 /*        body->atime = obj->atime;
 342         body->ctime = obj->ctime;
 343         body->mtime = obj->mtime;
 344         body->nlink = obj->nlink;*/
 345 }
 346
 347 int lmv_lookup_slaves(struct obd_export *exp, struct ptlrpc_request **reqp)
 348 {
 349         struct obd_device *obd = exp->exp_obd;
 350         struct lmv_obd *lmv = &obd->u.lmv;
 351         struct mds_body *body = NULL;
 352         struct lustre_handle *lockh;
 353         struct ldlm_lock *lock;
 354         struct mds_body *body2;
 355         struct lmv_obj *obj;
 356         int i, rc = 0;
 357         ENTRY;
 358
 359         LASSERT(reqp);
 360         LASSERT(*reqp);
 361
 362         /* master is locked. we'd like to take locks on slaves and update
 363          * attributes to be returned from the slaves it's important that lookup
 364          * is called in two cases:
 365
 366          *  - for first time (dcache has no such a resolving yet).
 367          *  - ->d_revalidate() returned false.
 368
 369          * last case possible only if all the objs (master and all slaves aren't
 370          * valid */
 371
 372         body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 373         LASSERT(body != NULL);
 374
 375         obj = lmv_grab_obj(obd, &body->fid1);
 376         LASSERT(obj != NULL);
 377
 378         CDEBUG(D_OTHER, "lookup slaves for %lu/%lu/%lu\n",
 379                (unsigned long)body->fid1.mds,
 380                (unsigned long)body->fid1.id,
 381                (unsigned long)body->fid1.generation);
 382
 383         lmv_lock_obj(obj);
 384
 385         for (i = 0; i < obj->objcount; i++) {
 386                 struct ll_fid fid = obj->objs[i].fid;
 387                 struct ptlrpc_request *req = NULL;
 388                 struct lookup_intent it;
 389
 390                 if (fid_equal(&fid, &obj->fid))
 391                         /* skip master obj */
 392                         continue;
 393
 394                 CDEBUG(D_OTHER, "lookup slave %lu/%lu/%lu\n",
 395                        (unsigned long)fid.mds, (unsigned long)fid.id,
 396                        (unsigned long)fid.generation);
 397
 398                 /* is obj valid? */
 399                 memset(&it, 0, sizeof(it));
 400                 it.it_op = IT_GETATTR;
 401                 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &fid,
 402                                     NULL, 0, NULL, 0, &fid, &it, 0, &req,
 403                                     lmv_dirobj_blocking_ast);
 404
 405                 lockh = (struct lustre_handle *)&it.d.lustre.it_lock_handle;
 406                 if (rc > 0) {
 407                         /* nice, this slave is valid */
 408                         LASSERT(req == NULL);
 409                         CDEBUG(D_OTHER, "cached\n");
 410                         goto release_lock;
 411                 }
 412
 413                 if (rc < 0)
 414                         /* error during revalidation */
 415                         GOTO(cleanup, rc);
 416
 417                 /* rc == 0, this means we have no such a lock and can't think
 418                  * obj is still valid. lookup it again */
 419                 LASSERT(req == NULL);
 420                 req = NULL;
 421
 422                 memset(&it, 0, sizeof(it));
 423                 it.it_op = IT_GETATTR;
 424                 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &fid,
 425                                     NULL, 0, NULL, 0, NULL, &it, 0, &req,
 426                                     lmv_dirobj_blocking_ast);
 427
 428                 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
 429                 LASSERT(rc <= 0);
 430
 431                 if (rc < 0)
 432                         /* error during lookup */
 433                         GOTO(cleanup, rc);
 434
 435                 lock = ldlm_handle2lock(lockh);
 436                 LASSERT(lock);
 437
 438                 lock->l_ast_data = lmv_get_obj(obj);
 439
 440                 body2 = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body2));
 441                 LASSERT(body2);
 442
 443                 obj->objs[i].size = body2->size;
 444
 445                 CDEBUG(D_OTHER, "fresh: %lu\n",
 446                        (unsigned long)obj->objs[i].size);
 447
 448                 LDLM_LOCK_PUT(lock);
 449
 450                 if (req)
 451                         ptlrpc_req_finished(req);
 452 release_lock:
 453                 lmv_update_body_from_obj(body, obj->objs + i);
 454
 455                 if (it.d.lustre.it_lock_mode)
 456                         ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
 457         }
 458 cleanup:
 459         lmv_unlock_obj(obj);
 460         lmv_put_obj(obj);
 461         RETURN(rc);
 462 }
 463
 464 int lmv_intent_lookup(struct obd_export *exp,
 465                       struct ll_fid *pfid, const char *name, int len,
 466                       void *lmm, int lmmsize, struct ll_fid *cfid,
 467                       struct lookup_intent *it, int flags,
 468                       struct ptlrpc_request **reqp,
 469                       ldlm_blocking_callback cb_blocking)
 470 {
 471         struct obd_device *obd = exp->exp_obd;
 472         struct lmv_obd *lmv = &obd->u.lmv;
 473         struct mds_body *body = NULL;
 474         struct ll_fid rpfid = *pfid;
 475         struct lmv_obj *obj;
 476         struct mea *mea;
 477         int rc, mds, loop = 0;
 478         ENTRY;
 479
 480         /* IT_LOOKUP is intended to produce name -> fid resolving (let's call
 481          * this lookup below) or to confirm requested resolving is still valid
 482          * (let's call this revalidation) cfid != NULL specifies revalidation */
 483
 484         if (cfid) {
 485                 /* this is revalidation: we have to check is LOOKUP lock still
 486                  * valid for given fid. very important part is that we have to
 487                  * choose right mds because namespace is per mds */
 488                 rpfid = *pfid;
 489                 obj = lmv_grab_obj(obd, pfid);
 490                 if (obj) {
 491                         mds = raw_name2idx(obj->hashtype, obj->objcount, (char *) name, len);
 492                         rpfid = obj->objs[mds].fid;
 493                         lmv_put_obj(obj);
 494                 }
 495                 mds = rpfid.mds;
 496
 497                 CDEBUG(D_OTHER, "revalidate lookup for %lu/%lu/%lu to %d MDS\n",
 498                        (unsigned long)cfid->mds, (unsigned long)cfid->id,
 499                        (unsigned long)cfid->generation, mds);
 500
 501                 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, pfid, name,
 502                                     len, lmm, lmmsize, cfid, it, flags,
 503                                     reqp, cb_blocking);
 504                 RETURN(rc);
 505         }
 506
 507         mds = pfid->mds;
 508 repeat:
 509         LASSERT(++loop <= 2);
 510         /* this is lookup. during lookup we have to update all the attributes,
 511          * because returned values will be put in struct inode */
 512
 513         obj = lmv_grab_obj(obd, pfid);
 514         if (obj) {
 515                 if (len) {
 516                         /* directory is already splitted. calculate mds */
 517                         mds = raw_name2idx(obj->hashtype, obj->objcount, (char *)name, len);
 518                         rpfid = obj->objs[mds].fid;
 519                         mds = rpfid.mds;
 520                 }
 521                 lmv_put_obj(obj);
 522         }
 523
 524         rc = md_intent_lock(lmv->tgts[mds].ltd_exp, &rpfid, name,
 525                             len, lmm, lmmsize, NULL, it, flags, reqp,
 526                             cb_blocking);
 527         if (rc > 0) {
 528                 /* very interesting. it seems object is still valid but for some
 529                  * reason llite calls lookup, not revalidate */
 530                 CWARN("lookup for %lu/%lu/%lu and data should be uptodate\n",
 531                       (unsigned long)rpfid.mds, (unsigned long)rpfid.id,
 532                       (unsigned long)rpfid.generation);
 533                 LASSERT(*reqp == NULL);
 534                 RETURN(rc);
 535         }
 536
 537         if (rc == 0 && *reqp == NULL) {
 538                 /* once again, we're asked for lookup, not revalidate */
 539                 CWARN("lookup for %lu/%lu/%lu and data should be uptodate\n",
 540                        (unsigned long)rpfid.mds, (unsigned long)rpfid.id,
 541                        (unsigned long)rpfid.generation);
 542                 RETURN(rc);
 543         }
 544
 545         if (rc == -ERESTART) {
 546                 /* directory got splitted since last update. this shouldn't be
 547                  * becasue splitting causes lock revocation, so revalidate had
 548                  * to fail and lookup on dir had to return mea */
 549                 CWARN("we haven't knew about directory splitting!\n");
 550                 LASSERT(obj == NULL);
 551
 552                 obj = lmv_create_obj(exp, &rpfid, NULL);
 553                 if (IS_ERR(obj))
 554                         RETURN(PTR_ERR(obj));
 555                 lmv_put_obj(obj);
 556                 goto repeat;
 557         }
 558
 559         if (rc < 0)
 560                 RETURN(rc);
 561
 562         /* okay, MDS has returned success. probably name has been resolved in
 563          * remote inode */
 564         rc = lmv_handle_remote_inode(exp, lmm, lmmsize, it, flags,
 565                                      reqp, cb_blocking);
 566
 567         if (rc == 0 && (mea = body_of_splitted_dir(*reqp, 1))) {
 568                 /* wow! this is splitted dir, we'd like to handle it */
 569                 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 570                 LASSERT(body != NULL);
 571
 572                 obj = lmv_grab_obj(obd, &body->fid1);
 573                 if (!obj) {
 574                         obj = lmv_create_obj(exp, &body->fid1, mea);
 575                         if (IS_ERR(obj))
 576                                 RETURN(PTR_ERR(obj));
 577                 }
 578                 lmv_put_obj(obj);
 579         }
 580
 581         RETURN(rc);
 582 }
 583
 584 int lmv_intent_lock(struct obd_export *exp,
 585                     struct ll_fid *pfid, const char *name, int len,
 586                     void *lmm, int lmmsize, struct ll_fid *cfid,
 587                     struct lookup_intent *it, int flags,
 588                     struct ptlrpc_request **reqp,
 589                     ldlm_blocking_callback cb_blocking)
 590 {
 591         struct obd_device *obd = exp->exp_obd;
 592         int rc = 0;
 593         ENTRY;
 594
 595         LASSERT(it);
 596         LASSERT(pfid);
 597
 598         CDEBUG(D_OTHER, "INTENT LOCK '%s' for '%*s' on %lu/%lu -> %u\n",
 599                         LL_IT2STR(it), len, name, (unsigned long) pfid->id,
 600                         (unsigned long) pfid->generation, pfid->mds);
 601
 602         rc = lmv_check_connect(obd);
 603         if (rc)
 604                 RETURN(rc);
 605
 606         if (it->it_op == IT_LOOKUP)
 607                 rc = lmv_intent_lookup(exp, pfid, name, len, lmm,
 608                                        lmmsize, cfid, it, flags, reqp,
 609                                        cb_blocking);
 610         else if (it->it_op & IT_OPEN)
 611                 rc = lmv_intent_open(exp, pfid, name, len, lmm,
 612                                      lmmsize, cfid, it, flags, reqp,
 613                                      cb_blocking);
 614         else if (it->it_op == IT_GETATTR || it->it_op == IT_CHDIR)
 615                 rc = lmv_intent_getattr(exp, pfid, name, len, lmm,
 616                                         lmmsize, cfid, it, flags, reqp,
 617                                         cb_blocking);
 618         else
 619                 LBUG();
 620         RETURN(rc);
 621 }
 622
 623 int lmv_revalidate_slaves(struct obd_export *exp, struct ptlrpc_request **reqp,
 624                           struct ll_fid *mfid, struct lookup_intent *oit,
 625                           int master_valid, ldlm_blocking_callback cb_blocking)
 626 {
 627         struct obd_device *obd = exp->exp_obd;
 628         struct ptlrpc_request *mreq = *reqp;
 629         struct lmv_obd *lmv = &obd->u.lmv;
 630         struct lustre_handle master_lockh;
 631         struct ldlm_lock *lock;
 632         unsigned long size = 0;
 633         struct mds_body *body;
 634         struct lmv_obj *obj;
 635         int master_lock_mode;
 636         int i, rc = 0;
 637         ENTRY;
 638
 639         /* we have to loop over the subobjects, check validity and update them
 640          * from MDSs if needed. it's very useful that we need not to update all
 641          * the fields. say, common fields (that are equal on all the subojects
 642          * need not to be update, another fields (i_size, for example) are
 643          * cached all the time */
 644         obj = lmv_grab_obj(obd, mfid);
 645         LASSERT(obj != NULL);
 646
 647         master_lock_mode = 0;
 648
 649         lmv_lock_obj(obj);
 650
 651         for (i = 0; i < obj->objcount; i++) {
 652                 struct ll_fid fid = obj->objs[i].fid;
 653                 struct lustre_handle *lockh = NULL;
 654                 struct ptlrpc_request *req = NULL;
 655                 ldlm_blocking_callback cb;
 656                 struct lookup_intent it;
 657                 int master = 0;
 658
 659                 CDEBUG(D_OTHER, "revalidate subobj %lu/%lu/%lu\n",
 660                        (unsigned long)fid.mds, (unsigned long)fid.id,
 661                        (unsigned long) fid.generation);
 662
 663                 memset(&it, 0, sizeof(it));
 664                 it.it_op = IT_GETATTR;
 665                 cb = lmv_dirobj_blocking_ast;
 666
 667                 if (fid_equal(&fid, &obj->fid)) {
 668                         if (master_valid) {
 669                                 /* lmv_intent_getattr() already checked
 670                                  * validness and took the lock */
 671                                 if (mreq) {
 672                                         /* it even got the reply refresh attrs
 673                                          * from that reply */
 674                                         body = lustre_msg_buf(mreq->rq_repmsg,
 675                                                               1, sizeof(*body));
 676                                         LASSERT(body != NULL);
 677                                         goto update;
 678                                 }
 679                                 /* take already cached attrs into account */
 680                                 CDEBUG(D_OTHER,
 681                                        "master is locked and cached\n");
 682                                 goto release_lock;
 683                         }
 684                         master = 1;
 685                         cb = cb_blocking;
 686                 }
 687
 688                 /* is obj valid? */
 689                 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &fid,
 690                                     NULL, 0, NULL, 0, &fid, &it, 0, &req, cb);
 691                 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
 692                 if (rc > 0) {
 693                         /* nice, this slave is valid */
 694                         LASSERT(req == NULL);
 695                         CDEBUG(D_OTHER, "cached\n");
 696                         goto release_lock;
 697                 }
 698
 699                 if (rc < 0)
 700                         /* error during revalidation */
 701                         GOTO(cleanup, rc);
 702
 703                 /* rc == 0, this means we have no such a lock and can't think
 704                  * obj is still valid. lookup it again */
 705                 LASSERT(req == NULL);
 706                 req = NULL;
 707
 708                 memset(&it, 0, sizeof(it));
 709                 it.it_op = IT_GETATTR;
 710                 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &fid,
 711                                     NULL, 0, NULL, 0, NULL, &it, 0, &req, cb);
 712                 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
 713                 LASSERT(rc <= 0);
 714
 715                 if (rc < 0)
 716                         /* error during lookup */
 717                         GOTO(cleanup, rc);
 718
 719                 if (master) {
 720                         LASSERT(master_valid == 0);
 721                         /* save lock on master to be returned to the caller */
 722                         CDEBUG(D_OTHER, "no lock on master yet\n");
 723                         memcpy(&master_lockh, lockh, sizeof(master_lockh));
 724                         master_lock_mode = it.d.lustre.it_lock_mode;
 725                         it.d.lustre.it_lock_mode = 0;
 726                 } else {
 727                         /* this is slave. we want to control it */
 728                         lock = ldlm_handle2lock(lockh);
 729                         LASSERT(lock);
 730                         lock->l_ast_data = lmv_get_obj(obj);
 731                         LDLM_LOCK_PUT(lock);
 732                 }
 733
 734                 if (*reqp == NULL) {
 735                         /* this is first reply, we'll use it to return
 736                          * updated data back to the caller */
 737                         LASSERT(req);
 738                         ptlrpc_request_addref(req);
 739                         *reqp = req;
 740
 741                 }
 742
 743                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
 744                 LASSERT(body);
 745
 746 update:
 747                 obj->objs[i].size = body->size;
 748
 749                 CDEBUG(D_OTHER, "fresh: %lu\n",
 750                        (unsigned long)obj->objs[i].size);
 751
 752                 if (req)
 753                         ptlrpc_req_finished(req);
 754 release_lock:
 755                 size += obj->objs[i].size;
 756
 757                 if (it.d.lustre.it_lock_mode)
 758                         ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
 759         }
 760
 761         if (*reqp) {
 762                 /* some attrs got refreshed, we have reply and it's time to put
 763                  * fresh attrs to it */
 764                 CDEBUG(D_OTHER, "return refreshed attrs: size = %lu\n",
 765                        (unsigned long)size);
 766
 767                 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 768                 LASSERT(body);
 769
 770                 /* FIXME: what about other attributes? */
 771                 body->size = size;
 772
 773                 if (mreq == NULL) {
 774                         /* very important to maintain lli->mds the same because
 775                          * of revalidation. mreq == NULL means that caller has
 776                          * no reply and the only attr we can return is size */
 777                         body->valid = OBD_MD_FLSIZE;
 778                         body->mds = obj->fid.mds;
 779                 }
 780                 if (master_valid == 0) {
 781                         memcpy(&oit->d.lustre.it_lock_handle,
 782                                &master_lockh, sizeof(master_lockh));
 783                         oit->d.lustre.it_lock_mode = master_lock_mode;
 784                 }
 785                 rc = 0;
 786         } else {
 787                 /* it seems all the attrs are fresh and we did no request */
 788                 CDEBUG(D_OTHER, "all the attrs were fresh\n");
 789                 if (master_valid == 0)
 790                         oit->d.lustre.it_lock_mode = master_lock_mode;
 791                 rc = 1;
 792         }
 793 cleanup:
 794         lmv_unlock_obj(obj);
 795         lmv_put_obj(obj);
 796         RETURN(rc);
 797 }