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