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