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, struct ll_uctxt *uctxt,
  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;
  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, uctxt, &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, struct ll_uctxt *uctxt,
 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;
 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         mds = rpfid.mds;
 141         obj = lmv_grab_obj(obd, &rpfid);
 142         if (obj) {
 143                 /* directory is already splitted, so we have to forward
 144                  * request to the right MDS */
 145                 mds = raw_name2idx(obj->objcount, (char *)name, len);
 146                 CDEBUG(D_OTHER, "forward to MDS #%u\n", mds);
 147
 148                 rpfid = obj->objs[mds].fid;
 149                 lmv_put_obj(obj);
 150         }
 151
 152         rc = md_intent_lock(lmv->tgts[mds].ltd_exp, uctxt, &rpfid, name,
 153                             len, lmm, lmmsize, cfid, it, flags, reqp,
 154                             cb_blocking);
 155         if (rc == -ERESTART) {
 156                 /* directory got splitted. time to update local object
 157                  * and repeat the request with proper MDS */
 158                 LASSERT(fid_equal(pfid, &rpfid));
 159                 rc = lmv_get_mea_and_update_object(exp, &rpfid);
 160                 if (rc == 0) {
 161                         ptlrpc_req_finished(*reqp);
 162                         goto repeat;
 163                 }
 164         }
 165         if (rc != 0)
 166                 RETURN(rc);
 167
 168         /* okay, MDS has returned success. Probably name has been resolved in
 169          * remote inode */
 170         rc = lmv_handle_remote_inode(exp, uctxt, lmm, lmmsize, it,
 171                                      flags, reqp, cb_blocking);
 172         if (rc != 0) {
 173                 LASSERT(rc < 0);
 174                 RETURN(rc);
 175         }
 176
 177         /* caller may use attrs MDS returns on IT_OPEN lock request so, we have
 178          * to update them for splitted dir */
 179         body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 180         LASSERT(body != NULL);
 181
 182         cfid = &body->fid1;
 183         obj = lmv_grab_obj(obd, cfid);
 184         if (!obj && (mea = body_of_splitted_dir(*reqp, 1))) {
 185                 /* wow! this is splitted dir, we'd like to handle it */
 186                 obj = lmv_create_obj(exp, &body->fid1, mea);
 187                 if (IS_ERR(obj))
 188                         RETURN(PTR_ERR(obj));
 189         }
 190
 191         if (obj) {
 192                 /* this is splitted dir and we'd want to get attrs */
 193                 CDEBUG(D_OTHER, "attrs from slaves for %lu/%lu/%lu\n",
 194                        (unsigned long)cfid->mds, (unsigned long)cfid->id,
 195                        (unsigned long)cfid->generation);
 196                 rc = lmv_revalidate_slaves(exp, reqp, cfid,
 197                                            it, 1, cb_blocking);
 198         } else if (S_ISDIR(body->mode)) {
 199                 /*CWARN("hmmm, %lu/%lu/%lu has not lmv obj?!\n",
 200                                 (unsigned long) cfid->mds,
 201                                 (unsigned long) cfid->id,
 202                                 (unsigned long) cfid->generation);*/
 203         }
 204
 205         if (obj)
 206                 lmv_put_obj(obj);
 207
 208         RETURN(rc);
 209 }
 210
 211 int lmv_intent_getattr(struct obd_export *exp, struct ll_uctxt *uctxt,
 212                        struct ll_fid *pfid, const char *name, int len,
 213                        void *lmm, int lmmsize, struct ll_fid *cfid,
 214                        struct lookup_intent *it, int flags,
 215                        struct ptlrpc_request **reqp,
 216                        ldlm_blocking_callback cb_blocking)
 217 {
 218         struct obd_device *obd = exp->exp_obd;
 219         struct lmv_obd *lmv = &obd->u.lmv;
 220         struct mds_body *body = NULL;
 221         struct ll_fid rpfid = *pfid;
 222         struct lmv_obj *obj, *obj2;
 223         struct mea *mea;
 224         int rc = 0, mds;
 225         ENTRY;
 226
 227         if (cfid) {
 228                 /* caller wants to revalidate attrs of obj we have to revalidate
 229                  * slaves if requested object is splitted directory */
 230                 CDEBUG(D_OTHER, "revalidate attrs for %lu/%lu/%lu\n",
 231                        (unsigned long)cfid->mds, (unsigned long)cfid->id,
 232                        (unsigned long)cfid->generation);
 233                 mds = cfid->mds;
 234                 obj = lmv_grab_obj(obd, cfid);
 235                 if (obj) {
 236                         /* in fact, we need not this with current intent_lock(),
 237                          * but it may change some day */
 238                         rpfid = obj->objs[mds].fid;
 239                         lmv_put_obj(obj);
 240                 }
 241                 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, uctxt, &rpfid, name,
 242                                     len, lmm, lmmsize, cfid, it, flags, reqp,
 243                                     cb_blocking);
 244                 if (obj && rc >= 0) {
 245                         /* this is splitted dir. In order to optimize things a
 246                          * bit, we consider obj valid updating missing parts.
 247
 248                          * FIXME: do we need to return any lock here? It would
 249                          * be fine if we don't. this means that nobody should
 250                          * use UPDATE lock to notify about object * removal */
 251                         CDEBUG(D_OTHER,
 252                                "revalidate slaves for %lu/%lu/%lu, rc %d\n",
 253                                (unsigned long)cfid->mds, (unsigned long)cfid->id,
 254                                (unsigned long)cfid->generation, rc);
 255
 256                         rc = lmv_revalidate_slaves(exp, reqp, cfid, it, rc,
 257                                                    cb_blocking);
 258                 }
 259
 260                 RETURN(rc);
 261         }
 262
 263         CDEBUG(D_OTHER, "INTENT getattr for %*s on %lu/%lu/%lu\n",
 264                len, name, (unsigned long)pfid->mds, (unsigned long)pfid->id,
 265                (unsigned long)pfid->generation);
 266
 267         mds = pfid->mds;
 268         obj = lmv_grab_obj(obd, pfid);
 269         if (obj && len) {
 270                 /* directory is already splitted. calculate mds */
 271                 mds = raw_name2idx(obj->objcount, (char *) name, len);
 272                 rpfid = obj->objs[mds].fid;
 273                 lmv_put_obj(obj);
 274
 275                 CDEBUG(D_OTHER, "forward to MDS #%u (slave %lu/%lu/%lu)\n",
 276                        mds, (unsigned long)rpfid.mds, (unsigned long)rpfid.id,
 277                        (unsigned long)rpfid.generation);
 278         }
 279
 280         rc = md_intent_lock(lmv->tgts[mds].ltd_exp, uctxt, &rpfid, name,
 281                             len, lmm, lmmsize, NULL, it, flags, reqp,
 282                             cb_blocking);
 283
 284         if (rc < 0)
 285                 RETURN(rc);
 286
 287         LASSERT(rc == 0);
 288
 289         /* okay, MDS has returned success. probably name has been
 290          * resolved in remote inode */
 291         rc = lmv_handle_remote_inode(exp, uctxt, lmm, lmmsize, it,
 292                                      flags, reqp, cb_blocking);
 293         if (rc < 0)
 294                 RETURN(rc);
 295
 296         body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 297         LASSERT(body != NULL);
 298
 299         cfid = &body->fid1;
 300         obj2 = lmv_grab_obj(obd, cfid);
 301
 302         if (!obj2 && (mea = body_of_splitted_dir(*reqp, 1))) {
 303                 /* wow! this is splitted dir, we'd like to handle it. */
 304                 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 305                 LASSERT(body != NULL);
 306
 307                 obj2 = lmv_create_obj(exp, &body->fid1, mea);
 308                 if (IS_ERR(obj2))
 309                         RETURN(PTR_ERR(obj2));
 310         }
 311
 312         if (obj2) {
 313                 /* this is splitted dir and we'd want to get attrs */
 314                 CDEBUG(D_OTHER, "attrs from slaves for %lu/%lu/%lu, rc %d\n",
 315                        (unsigned long)cfid->mds, (unsigned long)cfid->id,
 316                        (unsigned long)cfid->generation, rc);
 317
 318                 rc = lmv_revalidate_slaves(exp, reqp, cfid, it, 1, cb_blocking);
 319                 lmv_put_obj(obj2);
 320         }
 321         RETURN(rc);
 322 }
 323
 324 void lmv_update_body_from_obj(struct mds_body *body, struct lmv_inode *obj)
 325 {
 326         /* update size */
 327         body->size += obj->size;
 328
 329         /* update atime */
 330         /* update ctime */
 331         /* update mtime */
 332         /* update nlink */
 333 }
 334
 335 int lmv_lookup_slaves(struct obd_export *exp, struct ptlrpc_request **reqp)
 336 {
 337         struct obd_device *obd = exp->exp_obd;
 338         struct lmv_obd *lmv = &obd->u.lmv;
 339         struct mds_body *body = NULL;
 340         struct lustre_handle *lockh;
 341         struct ldlm_lock *lock;
 342         struct mds_body *body2;
 343         struct ll_uctxt uctxt;
 344         struct lmv_obj *obj;
 345         int i, rc = 0;
 346         ENTRY;
 347
 348         LASSERT(reqp);
 349         LASSERT(*reqp);
 350
 351         /* master is locked. we'd like to take locks on slaves and update
 352          * attributes to be returned from the slaves it's important that lookup
 353          * is called in two cases:
 354
 355          *  - for first time (dcache has no such a resolving yet.
 356          *  - ->d_revalidate() returned false.
 357
 358          * last case possible only if all the objs (master and all slaves aren't
 359          * valid */
 360
 361         body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 362         LASSERT(body != NULL);
 363
 364         obj = lmv_grab_obj(obd, &body->fid1);
 365         LASSERT(obj != NULL);
 366
 367         CDEBUG(D_OTHER, "lookup slaves for %lu/%lu/%lu\n",
 368                (unsigned long)body->fid1.mds,
 369                (unsigned long)body->fid1.id,
 370                (unsigned long)body->fid1.generation);
 371
 372         uctxt.gid1 = 0;
 373         uctxt.gid2 = 0;
 374
 375         lmv_lock_obj(obj);
 376
 377         for (i = 0; i < obj->objcount; i++) {
 378                 struct ll_fid fid = obj->objs[i].fid;
 379                 struct ptlrpc_request *req = NULL;
 380                 struct lookup_intent it;
 381
 382                 if (fid_equal(&fid, &obj->fid))
 383                         /* skip master obj */
 384                         continue;
 385
 386                 CDEBUG(D_OTHER, "lookup slave %lu/%lu/%lu\n",
 387                        (unsigned long)fid.mds, (unsigned long)fid.id,
 388                        (unsigned long)fid.generation);
 389
 390                 /* is obj valid? */
 391                 memset(&it, 0, sizeof(it));
 392                 it.it_op = IT_GETATTR;
 393                 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &uctxt, &fid,
 394                                     NULL, 0, NULL, 0, &fid, &it, 0, &req,
 395                                     lmv_dirobj_blocking_ast);
 396
 397                 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
 398                 if (rc > 0) {
 399                         /* nice, this slave is valid */
 400                         LASSERT(req == NULL);
 401                         CDEBUG(D_OTHER, "cached\n");
 402                         goto release_lock;
 403                 }
 404
 405                 if (rc < 0)
 406                         /* error during revalidation */
 407                         GOTO(cleanup, rc);
 408
 409                 /* rc == 0, this means we have no such a lock and can't think
 410                  * obj is still valid. lookup it again */
 411                 LASSERT(req == NULL);
 412                 req = NULL;
 413
 414                 memset(&it, 0, sizeof(it));
 415                 it.it_op = IT_GETATTR;
 416                 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &uctxt, &fid,
 417                                     NULL, 0, NULL, 0, NULL, &it, 0, &req,
 418                                     lmv_dirobj_blocking_ast);
 419
 420                 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
 421                 LASSERT(rc <= 0);
 422
 423                 if (rc < 0)
 424                         /* error during lookup */
 425                         GOTO(cleanup, rc);
 426
 427                 lock = ldlm_handle2lock(lockh);
 428                 LASSERT(lock);
 429
 430                 lock->l_ast_data = lmv_get_obj(obj);
 431
 432                 body2 = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body2));
 433                 LASSERT(body2);
 434
 435                 obj->objs[i].size = body2->size;
 436                 CDEBUG(D_OTHER, "fresh: %lu\n",
 437                        (unsigned long)obj->objs[i].size);
 438
 439                 LDLM_LOCK_PUT(lock);
 440
 441                 if (req)
 442                         ptlrpc_req_finished(req);
 443 release_lock:
 444                 lmv_update_body_from_obj(body, obj->objs + i);
 445
 446                 if (it.d.lustre.it_lock_mode)
 447                         ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
 448         }
 449 cleanup:
 450         lmv_unlock_obj(obj);
 451         lmv_put_obj(obj);
 452         RETURN(rc);
 453 }
 454
 455 int lmv_intent_lookup(struct obd_export *exp, struct ll_uctxt *uctxt,
 456                       struct ll_fid *pfid, const char *name, int len,
 457                       void *lmm, int lmmsize, struct ll_fid *cfid,
 458                       struct lookup_intent *it, int flags,
 459                       struct ptlrpc_request **reqp,
 460                       ldlm_blocking_callback cb_blocking)
 461 {
 462         struct obd_device *obd = exp->exp_obd;
 463         struct lmv_obd *lmv = &obd->u.lmv;
 464         struct mds_body *body = NULL;
 465         struct ll_fid rpfid = *pfid;
 466         struct lmv_obj *obj;
 467         struct mea *mea;
 468         int rc, mds;
 469         ENTRY;
 470
 471         /* IT_LOOKUP is intended to produce name -> fid resolving (let's call
 472          * this lookup below) or to confirm requested resolving is still valid
 473          * (let's call this revalidation) cfid != NULL specifies revalidation */
 474
 475         if (cfid) {
 476                 /* this is revalidation: we have to check is LOOKUP lock still
 477                  * valid for given fid. very important part is that we have to
 478                  * choose right mds because namespace is per mds */
 479                 rpfid = *pfid;
 480                 obj = lmv_grab_obj(obd, pfid);
 481                 if (obj) {
 482                         mds = raw_name2idx(obj->objcount, (char *) name, len);
 483                         rpfid = obj->objs[mds].fid;
 484                         lmv_put_obj(obj);
 485                 }
 486                 mds = rpfid.mds;
 487
 488                 CDEBUG(D_OTHER, "revalidate lookup for %lu/%lu/%lu to %d MDS\n",
 489                        (unsigned long)cfid->mds, (unsigned long)cfid->id,
 490                        (unsigned long)cfid->generation, mds);
 491
 492                 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, uctxt, pfid, name,
 493                                     len, lmm, lmmsize, cfid, it, flags,
 494                                     reqp, cb_blocking);
 495                 RETURN(rc);
 496         }
 497
 498         mds = pfid->mds;
 499 repeat:
 500         /* this is lookup. during lookup we have to update all the attributes,
 501          * because returned values will be put in struct inode */
 502
 503         obj = lmv_grab_obj(obd, pfid);
 504         if (obj && len) {
 505                 /* directory is already splitted. calculate mds */
 506                 mds = raw_name2idx(obj->objcount, (char *)name, len);
 507                 rpfid = obj->objs[mds].fid;
 508                 lmv_put_obj(obj);
 509         }
 510
 511         rc = md_intent_lock(lmv->tgts[mds].ltd_exp, uctxt, &rpfid, name,
 512                             len, lmm, lmmsize, NULL, it, flags, reqp,
 513                             cb_blocking);
 514         if (rc > 0) {
 515                 /* very interesting. it seems object is still valid but for some
 516                  * reason llite calls lookup, not revalidate */
 517                 CWARN("lookup for %lu/%lu/%lu and data should be uptodate\n",
 518                       (unsigned long)rpfid.mds, (unsigned long)rpfid.id,
 519                       (unsigned long)rpfid.generation);
 520
 521                 LASSERT(*reqp == NULL);
 522                 RETURN(rc);
 523         }
 524
 525         if (rc == 0 && *reqp == NULL) {
 526                 /* once again, we're asked for lookup, not revalidate */
 527                 CWARN("lookup for %lu/%lu/%lu and data should be uptodate\n",
 528                        (unsigned long)rpfid.mds, (unsigned long)rpfid.id,
 529                        (unsigned long)rpfid.generation);
 530                 RETURN(rc);
 531         }
 532
 533         if (rc == -ERESTART) {
 534                 /* directory got splitted since last update. this shouldn't be
 535                  * becasue splitting causes lock revocation, so revalidate had
 536                  * to fail and lookup on dir had to return mea */
 537                 CWARN("we haven't knew about directory splitting!\n");
 538                 LASSERT(obj == NULL);
 539
 540                 obj = lmv_create_obj(exp, &rpfid, NULL);
 541                 if (IS_ERR(obj))
 542                         RETURN(PTR_ERR(obj));
 543
 544                 goto repeat;
 545         }
 546
 547         if (rc < 0)
 548                 RETURN(rc);
 549
 550         /* okay, MDS has returned success. probably name has been resolved in
 551          * remote inode */
 552         rc = lmv_handle_remote_inode(exp, uctxt, lmm, lmmsize, it, flags,
 553                                      reqp, cb_blocking);
 554
 555         if (rc == 0 && (mea = body_of_splitted_dir(*reqp, 1))) {
 556                 /* wow! this is splitted dir, we'd like to handle it */
 557                 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 558                 LASSERT(body != NULL);
 559
 560                 obj = lmv_grab_obj(obd, &body->fid1);
 561                 if (!obj) {
 562                         obj = lmv_create_obj(exp, &body->fid1, mea);
 563                         if (IS_ERR(obj))
 564                                 RETURN(PTR_ERR(obj));
 565                 } else {
 566                         lmv_put_obj(obj);
 567                 }
 568         }
 569
 570         RETURN(rc);
 571 }
 572
 573 int lmv_intent_lock(struct obd_export *exp, struct ll_uctxt *uctxt,
 574                     struct ll_fid *pfid, const char *name, int len,
 575                     void *lmm, int lmmsize, struct ll_fid *cfid,
 576                     struct lookup_intent *it, int flags,
 577                     struct ptlrpc_request **reqp,
 578                     ldlm_blocking_callback cb_blocking)
 579 {
 580         struct obd_device *obd = exp->exp_obd;
 581         int rc = 0;
 582         ENTRY;
 583
 584         LASSERT(it);
 585         LASSERT(pfid);
 586
 587         CDEBUG(D_OTHER, "INTENT LOCK '%s' for '%*s' on %lu/%lu -> %u\n",
 588                         LL_IT2STR(it), len, name, (unsigned long) pfid->id,
 589                         (unsigned long) pfid->generation, pfid->mds);
 590
 591         rc = lmv_check_connect(obd);
 592         if (rc)
 593                 RETURN(rc);
 594
 595         if (it->it_op == IT_LOOKUP)
 596                 rc = lmv_intent_lookup(exp, uctxt, pfid, name, len, lmm,
 597                                        lmmsize, cfid, it, flags, reqp,
 598                                        cb_blocking);
 599         else if (it->it_op & IT_OPEN)
 600                 rc = lmv_intent_open(exp, uctxt, pfid, name, len, lmm,
 601                                      lmmsize, cfid, it, flags, reqp,
 602                                      cb_blocking);
 603         else if (it->it_op == IT_GETATTR || it->it_op == IT_CHDIR)
 604                 rc = lmv_intent_getattr(exp, uctxt, pfid, name, len, lmm,
 605                                         lmmsize, cfid, it, flags, reqp,
 606                                         cb_blocking);
 607         else
 608                 LBUG();
 609         RETURN(rc);
 610 }
 611
 612 int lmv_revalidate_slaves(struct obd_export *exp, struct ptlrpc_request **reqp,
 613                           struct ll_fid *mfid, struct lookup_intent *oit,
 614                           int master_valid, ldlm_blocking_callback cb_blocking)
 615 {
 616         struct obd_device *obd = exp->exp_obd;
 617         struct ptlrpc_request *mreq = *reqp;
 618         struct lmv_obd *lmv = &obd->u.lmv;
 619         struct lustre_handle master_lockh;
 620         unsigned long size = 0;
 621         struct ldlm_lock *lock;
 622         struct mds_body *body;
 623         struct ll_uctxt uctxt;
 624         struct lmv_obj *obj;
 625         int master_lock_mode;
 626         int i, rc = 0;
 627         ENTRY;
 628
 629         /* we have to loop over the subobjects, check validity and update them
 630          * from MDSs if needed. it's very useful that we need not to update all
 631          * the fields. say, common fields (that are equal on all the subojects
 632          * need not to be update, another fields (i_size, for example) are
 633          * cached all the time */
 634         obj = lmv_grab_obj(obd, mfid);
 635         LASSERT(obj != NULL);
 636
 637         uctxt.gid1 = 0;
 638         uctxt.gid2 = 0;
 639         master_lock_mode = 0;
 640
 641         lmv_lock_obj(obj);
 642
 643         for (i = 0; i < obj->objcount; i++) {
 644                 struct ll_fid fid = obj->objs[i].fid;
 645                 struct lustre_handle *lockh = NULL;
 646                 struct ptlrpc_request *req = NULL;
 647                 ldlm_blocking_callback cb;
 648                 struct lookup_intent it;
 649                 int master = 0;
 650
 651                 CDEBUG(D_OTHER, "revalidate subobj %lu/%lu/%lu\n",
 652                        (unsigned long)fid.mds, (unsigned long)fid.id,
 653                        (unsigned long) fid.generation);
 654
 655                 memset(&it, 0, sizeof(it));
 656                 it.it_op = IT_GETATTR;
 657                 cb = lmv_dirobj_blocking_ast;
 658
 659                 if (fid_equal(&fid, &obj->fid)) {
 660                         if (master_valid) {
 661                                 /* lmv_intent_getattr() already checked
 662                                  * validness and took the lock */
 663                                 if (mreq) {
 664                                         /* it even got the reply refresh attrs
 665                                          * from that reply */
 666                                         body = lustre_msg_buf(mreq->rq_repmsg,
 667                                                               1,sizeof(*body));
 668                                         LASSERT(body != NULL);
 669                                         goto update;
 670                                 }
 671                                 /* take already cached attrs into account */
 672                                 CDEBUG(D_OTHER,
 673                                        "master is locked and cached\n");
 674                                 goto release_lock;
 675                         }
 676                         master = 1;
 677                         cb = cb_blocking;
 678                 }
 679
 680                 /* is obj valid? */
 681                 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &uctxt, &fid,
 682                                     NULL, 0, NULL, 0, &fid, &it, 0, &req, cb);
 683                 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
 684                 if (rc > 0) {
 685                         /* nice, this slave is valid */
 686                         LASSERT(req == NULL);
 687                         CDEBUG(D_OTHER, "cached\n");
 688                         goto release_lock;
 689                 }
 690
 691                 if (rc < 0)
 692                         /* error during revalidation */
 693                         GOTO(cleanup, rc);
 694
 695                 /* rc == 0, this means we have no such a lock and can't think
 696                  * obj is still valid. lookup it again */
 697                 LASSERT(req == NULL);
 698                 req = NULL;
 699
 700                 memset(&it, 0, sizeof(it));
 701                 it.it_op = IT_GETATTR;
 702                 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &uctxt, &fid,
 703                                     NULL, 0, NULL, 0, NULL, &it, 0, &req, cb);
 704                 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
 705                 LASSERT(rc <= 0);
 706
 707                 if (rc < 0)
 708                         /* error during lookup */
 709                         GOTO(cleanup, rc);
 710
 711                 if (master) {
 712                         LASSERT(master_valid == 0);
 713                         /* save lock on master to be returned to the caller */
 714                         CDEBUG(D_OTHER, "no lock on master yet\n");
 715                         memcpy(&master_lockh, lockh, sizeof(master_lockh));
 716                         master_lock_mode = it.d.lustre.it_lock_mode;
 717                         it.d.lustre.it_lock_mode = 0;
 718                 } else {
 719                         /* this is slave. we want to control it */
 720                         lock = ldlm_handle2lock(lockh);
 721                         LASSERT(lock);
 722                         lock->l_ast_data = lmv_get_obj(obj);
 723                         LDLM_LOCK_PUT(lock);
 724                 }
 725
 726                 if (*reqp == NULL) {
 727                         /* this is first reply, we'll use it to return
 728                          * updated data back to the caller */
 729                         LASSERT(req);
 730                         ptlrpc_request_addref(req);
 731                         *reqp = req;
 732
 733                 }
 734
 735                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
 736                 LASSERT(body);
 737
 738 update:
 739                 obj->objs[i].size = body->size;
 740                 CDEBUG(D_OTHER, "fresh: %lu\n",
 741                        (unsigned long)obj->objs[i].size);
 742
 743                 if (req)
 744                         ptlrpc_req_finished(req);
 745 release_lock:
 746                 size += obj->objs[i].size;
 747                 if (it.d.lustre.it_lock_mode)
 748                         ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
 749         }
 750
 751         if (*reqp) {
 752                 /* some attrs got refreshed, we have reply and it's time to put
 753                  * fresh attrs to it */
 754                 CDEBUG(D_OTHER, "return refreshed attrs: size = %lu\n",
 755                        (unsigned long) size);
 756                 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
 757                 LASSERT(body);
 758
 759                 /* FIXME: what about another attributes? */
 760                 body->size = size;
 761                 if (mreq == NULL) {
 762                         /* very important to maintain lli->mds the same because
 763                          * of revalidation. mreq == NULL means that caller has
 764                          * no reply and the only attr we can return is size */
 765                         body->valid = OBD_MD_FLSIZE;
 766                         body->mds = obj->fid.mds;
 767                 }
 768                 if (master_valid == 0) {
 769                         memcpy(&oit->d.lustre.it_lock_handle,
 770                                &master_lockh, sizeof(master_lockh));
 771                         oit->d.lustre.it_lock_mode = master_lock_mode;
 772                 }
 773                 rc = 0;
 774         } else {
 775                 /* it seems all the attrs are fresh and we did no request */
 776                 CDEBUG(D_OTHER, "all the attrs were fresh\n");
 777                 if (master_valid == 0)
 778                         oit->d.lustre.it_lock_mode = master_lock_mode;
 779                 rc = 1;
 780         }
 781 cleanup:
 782         lmv_unlock_obj(obj);
 783         lmv_put_obj(obj);
 784         RETURN(rc);
 785 }