lustre/lmv/lmv_intent.c

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