lustre/lov/lov_request.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LOV
  41
  42 #ifdef __KERNEL__
  43 #include <libcfs/libcfs.h>
  44 #else
  45 #include <liblustre.h>
  46 #endif
  47
  48 #include <obd_class.h>
  49 #include <obd_lov.h>
  50 #include <lustre/lustre_idl.h>
  51
  52 #include "lov_internal.h"
  53
  54 static void lov_init_set(struct lov_request_set *set)
  55 {
  56         set->set_count = 0;
  57         set->set_completes = 0;
  58         set->set_success = 0;
  59         set->set_cookies = 0;
  60         CFS_INIT_LIST_HEAD(&set->set_list);
  61         atomic_set(&set->set_refcount, 1);
  62 }
  63
  64 static void lov_finish_set(struct lov_request_set *set)
  65 {
  66         struct list_head *pos, *n;
  67         ENTRY;
  68
  69         LASSERT(set);
  70         list_for_each_safe(pos, n, &set->set_list) {
  71                 struct lov_request *req = list_entry(pos, struct lov_request,
  72                                                      rq_link);
  73                 list_del_init(&req->rq_link);
  74
  75                 if (req->rq_oi.oi_oa)
  76                         OBDO_FREE(req->rq_oi.oi_oa);
  77                 if (req->rq_oi.oi_md)
  78                         OBD_FREE(req->rq_oi.oi_md, req->rq_buflen);
  79                 if (req->rq_oi.oi_osfs)
  80                         OBD_FREE(req->rq_oi.oi_osfs,
  81                                  sizeof(*req->rq_oi.oi_osfs));
  82                 OBD_FREE(req, sizeof(*req));
  83         }
  84
  85         if (set->set_pga) {
  86                 int len = set->set_oabufs * sizeof(*set->set_pga);
  87                 OBD_FREE(set->set_pga, len);
  88         }
  89         if (set->set_lockh)
  90                 lov_llh_put(set->set_lockh);
  91
  92         OBD_FREE(set, sizeof(*set));
  93         EXIT;
  94 }
  95
  96 void lov_update_set(struct lov_request_set *set,
  97                     struct lov_request *req, int rc)
  98 {
  99         req->rq_complete = 1;
 100         req->rq_rc = rc;
 101
 102         set->set_completes++;
 103         if (rc == 0)
 104                 set->set_success++;
 105 }
 106
 107 int lov_update_common_set(struct lov_request_set *set,
 108                           struct lov_request *req, int rc)
 109 {
 110         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 111         ENTRY;
 112
 113         lov_update_set(set, req, rc);
 114
 115         /* grace error on inactive ost */
 116         if (rc && !(lov->lov_tgts[req->rq_idx] &&
 117                     lov->lov_tgts[req->rq_idx]->ltd_active))
 118                 rc = 0;
 119
 120         /* FIXME in raid1 regime, should return 0 */
 121         RETURN(rc);
 122 }
 123
 124 void lov_set_add_req(struct lov_request *req, struct lov_request_set *set)
 125 {
 126         list_add_tail(&req->rq_link, &set->set_list);
 127         set->set_count++;
 128 }
 129
 130 extern void osc_update_enqueue(struct lustre_handle *lov_lockhp,
 131                                struct lov_oinfo *loi, int flags,
 132                                struct ost_lvb *lvb, __u32 mode, int rc);
 133
 134 static int lov_update_enqueue_lov(struct obd_export *exp,
 135                                   struct lustre_handle *lov_lockhp,
 136                                   struct lov_oinfo *loi, int flags, int idx,
 137                                   __u64 oid, int rc)
 138 {
 139         struct lov_obd *lov = &exp->exp_obd->u.lov;
 140
 141         if (rc != ELDLM_OK &&
 142             !(rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT))) {
 143                 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
 144                 if (lov->lov_tgts[idx] && lov->lov_tgts[idx]->ltd_active) {
 145                         /* -EUSERS used by OST to report file contention */
 146                         if (rc != -EINTR && rc != -EUSERS)
 147                                 CERROR("enqueue objid "LPX64" subobj "
 148                                        LPX64" on OST idx %d: rc %d\n",
 149                                        oid, loi->loi_id, loi->loi_ost_idx, rc);
 150                 } else
 151                         rc = ELDLM_OK;
 152         }
 153         return rc;
 154 }
 155
 156 int lov_update_enqueue_set(struct lov_request *req, __u32 mode, int rc)
 157 {
 158         struct lov_request_set *set = req->rq_rqset;
 159         struct lustre_handle *lov_lockhp;
 160         struct obd_info *oi = set->set_oi;
 161         struct lov_oinfo *loi;
 162         ENTRY;
 163
 164         LASSERT(oi != NULL);
 165
 166         lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
 167         loi = oi->oi_md->lsm_oinfo[req->rq_stripe];
 168
 169         /* XXX LOV STACKING: OSC gets a copy, created in lov_prep_enqueue_set
 170          * and that copy can be arbitrarily out of date.
 171          *
 172          * The LOV API is due for a serious rewriting anyways, and this
 173          * can be addressed then. */
 174
 175         lov_stripe_lock(oi->oi_md);
 176         osc_update_enqueue(lov_lockhp, loi, oi->oi_flags,
 177                            &req->rq_oi.oi_md->lsm_oinfo[0]->loi_lvb, mode, rc);
 178         if (rc == ELDLM_LOCK_ABORTED && (oi->oi_flags & LDLM_FL_HAS_INTENT))
 179                 memset(lov_lockhp, 0, sizeof *lov_lockhp);
 180         rc = lov_update_enqueue_lov(set->set_exp, lov_lockhp, loi, oi->oi_flags,
 181                                     req->rq_idx, oi->oi_md->lsm_object_id, rc);
 182         lov_stripe_unlock(oi->oi_md);
 183         lov_update_set(set, req, rc);
 184         RETURN(rc);
 185 }
 186
 187 /* The callback for osc_enqueue that updates lov info for every OSC request. */
 188 static int cb_update_enqueue(void *cookie, int rc)
 189 {
 190         struct obd_info *oinfo = cookie;
 191         struct ldlm_enqueue_info *einfo;
 192         struct lov_request *lovreq;
 193
 194         lovreq = container_of(oinfo, struct lov_request, rq_oi);
 195         einfo = lovreq->rq_rqset->set_ei;
 196         return lov_update_enqueue_set(lovreq, einfo->ei_mode, rc);
 197 }
 198
 199 static int enqueue_done(struct lov_request_set *set, __u32 mode)
 200 {
 201         struct lov_request *req;
 202         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 203         int rc = 0;
 204         ENTRY;
 205
 206         /* enqueue/match success, just return */
 207         if (set->set_completes && set->set_completes == set->set_success)
 208                 RETURN(0);
 209
 210         /* cancel enqueued/matched locks */
 211         list_for_each_entry(req, &set->set_list, rq_link) {
 212                 struct lustre_handle *lov_lockhp;
 213
 214                 if (!req->rq_complete || req->rq_rc)
 215                         continue;
 216
 217                 lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
 218                 LASSERT(lov_lockhp);
 219                 if (!lustre_handle_is_used(lov_lockhp))
 220                         continue;
 221
 222                 rc = obd_cancel(lov->lov_tgts[req->rq_idx]->ltd_exp,
 223                                 req->rq_oi.oi_md, mode, lov_lockhp);
 224                 if (rc && lov->lov_tgts[req->rq_idx] &&
 225                     lov->lov_tgts[req->rq_idx]->ltd_active)
 226                         CERROR("cancelling obdjid "LPX64" on OST "
 227                                "idx %d error: rc = %d\n",
 228                                req->rq_oi.oi_md->lsm_object_id,
 229                                req->rq_idx, rc);
 230         }
 231         if (set->set_lockh)
 232                 lov_llh_put(set->set_lockh);
 233         RETURN(rc);
 234 }
 235
 236 int lov_fini_enqueue_set(struct lov_request_set *set, __u32 mode, int rc,
 237                          struct ptlrpc_request_set *rqset)
 238 {
 239         int ret = 0;
 240         ENTRY;
 241
 242         if (set == NULL)
 243                 RETURN(0);
 244         LASSERT(set->set_exp);
 245         /* Do enqueue_done only for sync requests and if any request
 246          * succeeded. */
 247         if (!rqset) {
 248                 if (rc)
 249                         set->set_completes = 0;
 250                 ret = enqueue_done(set, mode);
 251         } else if (set->set_lockh)
 252                 lov_llh_put(set->set_lockh);
 253
 254         if (atomic_dec_and_test(&set->set_refcount))
 255                 lov_finish_set(set);
 256
 257         RETURN(rc ? rc : ret);
 258 }
 259
 260 int lov_prep_enqueue_set(struct obd_export *exp, struct obd_info *oinfo,
 261                          struct ldlm_enqueue_info *einfo,
 262                          struct lov_request_set **reqset)
 263 {
 264         struct lov_obd *lov = &exp->exp_obd->u.lov;
 265         struct lov_request_set *set;
 266         int i, rc = 0;
 267         ENTRY;
 268
 269         OBD_ALLOC(set, sizeof(*set));
 270         if (set == NULL)
 271                 RETURN(-ENOMEM);
 272         lov_init_set(set);
 273
 274         set->set_exp = exp;
 275         set->set_oi = oinfo;
 276         set->set_ei = einfo;
 277         set->set_lockh = lov_llh_new(oinfo->oi_md);
 278         if (set->set_lockh == NULL)
 279                 GOTO(out_set, rc = -ENOMEM);
 280         oinfo->oi_lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 281
 282         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
 283                 struct lov_oinfo *loi;
 284                 struct lov_request *req;
 285                 obd_off start, end;
 286
 287                 loi = oinfo->oi_md->lsm_oinfo[i];
 288                 if (!lov_stripe_intersects(oinfo->oi_md, i,
 289                                            oinfo->oi_policy.l_extent.start,
 290                                            oinfo->oi_policy.l_extent.end,
 291                                            &start, &end))
 292                         continue;
 293
 294                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 295                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 296                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 297                         continue;
 298                 }
 299
 300                 OBD_ALLOC(req, sizeof(*req));
 301                 if (req == NULL)
 302                         GOTO(out_set, rc = -ENOMEM);
 303
 304                 req->rq_buflen = sizeof(*req->rq_oi.oi_md) +
 305                         sizeof(struct lov_oinfo *) +
 306                         sizeof(struct lov_oinfo);
 307                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
 308                 if (req->rq_oi.oi_md == NULL) {
 309                         OBD_FREE(req, sizeof(*req));
 310                         GOTO(out_set, rc = -ENOMEM);
 311                 }
 312                 req->rq_oi.oi_md->lsm_oinfo[0] =
 313                         ((void *)req->rq_oi.oi_md) + sizeof(*req->rq_oi.oi_md) +
 314                         sizeof(struct lov_oinfo *);
 315
 316
 317                 req->rq_rqset = set;
 318                 /* Set lov request specific parameters. */
 319                 req->rq_oi.oi_lockh = set->set_lockh->llh_handles + i;
 320                 req->rq_oi.oi_cb_up = cb_update_enqueue;
 321                 req->rq_oi.oi_flags = oinfo->oi_flags;
 322
 323                 LASSERT(req->rq_oi.oi_lockh);
 324
 325                 req->rq_oi.oi_policy.l_extent.gid =
 326                         oinfo->oi_policy.l_extent.gid;
 327                 req->rq_oi.oi_policy.l_extent.start = start;
 328                 req->rq_oi.oi_policy.l_extent.end = end;
 329
 330                 req->rq_idx = loi->loi_ost_idx;
 331                 req->rq_stripe = i;
 332
 333                 /* XXX LOV STACKING: submd should be from the subobj */
 334                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 335                 req->rq_oi.oi_md->lsm_object_gr = oinfo->oi_md->lsm_object_gr;
 336                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 337                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_kms_valid =
 338                         loi->loi_kms_valid;
 339                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_kms = loi->loi_kms;
 340                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_lvb = loi->loi_lvb;
 341
 342                 lov_set_add_req(req, set);
 343         }
 344         if (!set->set_count)
 345                 GOTO(out_set, rc = -EIO);
 346         *reqset = set;
 347         RETURN(0);
 348 out_set:
 349         lov_fini_enqueue_set(set, einfo->ei_mode, rc, NULL);
 350         RETURN(rc);
 351 }
 352
 353 int lov_update_match_set(struct lov_request_set *set, struct lov_request *req,
 354                          int rc)
 355 {
 356         int ret = rc;
 357         ENTRY;
 358
 359         if (rc > 0)
 360                 ret = 0;
 361         else if (rc == 0)
 362                 ret = 1;
 363         lov_update_set(set, req, ret);
 364         RETURN(rc);
 365 }
 366
 367 int lov_fini_match_set(struct lov_request_set *set, __u32 mode, int flags)
 368 {
 369         int rc = 0;
 370         ENTRY;
 371
 372         if (set == NULL)
 373                 RETURN(0);
 374         LASSERT(set->set_exp);
 375         rc = enqueue_done(set, mode);
 376         if ((set->set_count == set->set_success) &&
 377             (flags & LDLM_FL_TEST_LOCK))
 378                 lov_llh_put(set->set_lockh);
 379
 380         if (atomic_dec_and_test(&set->set_refcount))
 381                 lov_finish_set(set);
 382
 383         RETURN(rc);
 384 }
 385
 386 int lov_prep_match_set(struct obd_export *exp, struct obd_info *oinfo,
 387                        struct lov_stripe_md *lsm, ldlm_policy_data_t *policy,
 388                        __u32 mode, struct lustre_handle *lockh,
 389                        struct lov_request_set **reqset)
 390 {
 391         struct lov_obd *lov = &exp->exp_obd->u.lov;
 392         struct lov_request_set *set;
 393         int i, rc = 0;
 394         ENTRY;
 395
 396         OBD_ALLOC(set, sizeof(*set));
 397         if (set == NULL)
 398                 RETURN(-ENOMEM);
 399         lov_init_set(set);
 400
 401         set->set_exp = exp;
 402         set->set_oi = oinfo;
 403         set->set_oi->oi_md = lsm;
 404         set->set_lockh = lov_llh_new(lsm);
 405         if (set->set_lockh == NULL)
 406                 GOTO(out_set, rc = -ENOMEM);
 407         lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 408
 409         for (i = 0; i < lsm->lsm_stripe_count; i++){
 410                 struct lov_oinfo *loi;
 411                 struct lov_request *req;
 412                 obd_off start, end;
 413
 414                 loi = lsm->lsm_oinfo[i];
 415                 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
 416                                            policy->l_extent.end, &start, &end))
 417                         continue;
 418
 419                 /* FIXME raid1 should grace this error */
 420                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 421                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 422                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 423                         GOTO(out_set, rc = -EIO);
 424                 }
 425
 426                 OBD_ALLOC(req, sizeof(*req));
 427                 if (req == NULL)
 428                         GOTO(out_set, rc = -ENOMEM);
 429
 430                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 431                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
 432                 if (req->rq_oi.oi_md == NULL) {
 433                         OBD_FREE(req, sizeof(*req));
 434                         GOTO(out_set, rc = -ENOMEM);
 435                 }
 436
 437                 req->rq_oi.oi_policy.l_extent.start = start;
 438                 req->rq_oi.oi_policy.l_extent.end = end;
 439                 req->rq_oi.oi_policy.l_extent.gid = policy->l_extent.gid;
 440
 441                 req->rq_idx = loi->loi_ost_idx;
 442                 req->rq_stripe = i;
 443
 444                 /* XXX LOV STACKING: submd should be from the subobj */
 445                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 446                 req->rq_oi.oi_md->lsm_object_gr = lsm->lsm_object_gr;
 447                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 448
 449                 lov_set_add_req(req, set);
 450         }
 451         if (!set->set_count)
 452                 GOTO(out_set, rc = -EIO);
 453         *reqset = set;
 454         RETURN(rc);
 455 out_set:
 456         lov_fini_match_set(set, mode, 0);
 457         RETURN(rc);
 458 }
 459
 460 int lov_fini_cancel_set(struct lov_request_set *set)
 461 {
 462         int rc = 0;
 463         ENTRY;
 464
 465         if (set == NULL)
 466                 RETURN(0);
 467
 468         LASSERT(set->set_exp);
 469         if (set->set_lockh)
 470                 lov_llh_put(set->set_lockh);
 471
 472         if (atomic_dec_and_test(&set->set_refcount))
 473                 lov_finish_set(set);
 474
 475         RETURN(rc);
 476 }
 477
 478 int lov_prep_cancel_set(struct obd_export *exp, struct obd_info *oinfo,
 479                         struct lov_stripe_md *lsm, __u32 mode,
 480                         struct lustre_handle *lockh,
 481                         struct lov_request_set **reqset)
 482 {
 483         struct lov_request_set *set;
 484         int i, rc = 0;
 485         ENTRY;
 486
 487         OBD_ALLOC(set, sizeof(*set));
 488         if (set == NULL)
 489                 RETURN(-ENOMEM);
 490         lov_init_set(set);
 491
 492         set->set_exp = exp;
 493         set->set_oi = oinfo;
 494         set->set_oi->oi_md = lsm;
 495         set->set_lockh = lov_handle2llh(lockh);
 496         if (set->set_lockh == NULL) {
 497                 CERROR("LOV: invalid lov lock handle %p\n", lockh);
 498                 GOTO(out_set, rc = -EINVAL);
 499         }
 500         lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 501
 502         for (i = 0; i < lsm->lsm_stripe_count; i++){
 503                 struct lov_request *req;
 504                 struct lustre_handle *lov_lockhp;
 505                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 506
 507                 lov_lockhp = set->set_lockh->llh_handles + i;
 508                 if (!lustre_handle_is_used(lov_lockhp)) {
 509                         CDEBUG(D_RPCTRACE,"lov idx %d subobj "LPX64" no lock\n",
 510                                loi->loi_ost_idx, loi->loi_id);
 511                         continue;
 512                 }
 513
 514                 OBD_ALLOC(req, sizeof(*req));
 515                 if (req == NULL)
 516                         GOTO(out_set, rc = -ENOMEM);
 517
 518                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 519                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
 520                 if (req->rq_oi.oi_md == NULL) {
 521                         OBD_FREE(req, sizeof(*req));
 522                         GOTO(out_set, rc = -ENOMEM);
 523                 }
 524
 525                 req->rq_idx = loi->loi_ost_idx;
 526                 req->rq_stripe = i;
 527
 528                 /* XXX LOV STACKING: submd should be from the subobj */
 529                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 530                 req->rq_oi.oi_md->lsm_object_gr = lsm->lsm_object_gr;
 531                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 532
 533                 lov_set_add_req(req, set);
 534         }
 535         if (!set->set_count)
 536                 GOTO(out_set, rc = -EIO);
 537         *reqset = set;
 538         RETURN(rc);
 539 out_set:
 540         lov_fini_cancel_set(set);
 541         RETURN(rc);
 542 }
 543
 544 static int create_done(struct obd_export *exp, struct lov_request_set *set,
 545                        struct lov_stripe_md **lsmp)
 546 {
 547         struct lov_obd *lov = &exp->exp_obd->u.lov;
 548         struct obd_trans_info *oti = set->set_oti;
 549         struct obdo *src_oa = set->set_oi->oi_oa;
 550         struct lov_request *req;
 551         struct obdo *ret_oa = NULL;
 552         int attrset = 0, rc = 0;
 553         ENTRY;
 554
 555         LASSERT(set->set_completes);
 556
 557         /* try alloc objects on other osts if osc_create fails for
 558          * exceptions: RPC failure, ENOSPC, etc */
 559         if (set->set_count != set->set_success) {
 560                 list_for_each_entry (req, &set->set_list, rq_link) {
 561                         if (req->rq_rc == 0)
 562                                 continue;
 563
 564                         set->set_completes--;
 565                         req->rq_complete = 0;
 566
 567                         rc = qos_remedy_create(set, req);
 568                         lov_update_create_set(set, req, rc);
 569
 570                         if (rc)
 571                                 break;
 572                 }
 573         }
 574
 575         /* no successful creates */
 576         if (set->set_success == 0)
 577                 GOTO(cleanup, rc);
 578
 579         /* If there was an explicit stripe set, fail.  Otherwise, we
 580          * got some objects and that's not bad. */
 581         if (set->set_count != set->set_success) {
 582                 if (*lsmp)
 583                         GOTO(cleanup, rc);
 584                 set->set_count = set->set_success;
 585                 qos_shrink_lsm(set);
 586         }
 587
 588         OBDO_ALLOC(ret_oa);
 589         if (ret_oa == NULL)
 590                 GOTO(cleanup, rc = -ENOMEM);
 591
 592         list_for_each_entry(req, &set->set_list, rq_link) {
 593                 if (!req->rq_complete || req->rq_rc)
 594                         continue;
 595                 lov_merge_attrs(ret_oa, req->rq_oi.oi_oa,
 596                                 req->rq_oi.oi_oa->o_valid, set->set_oi->oi_md,
 597                                 req->rq_stripe, &attrset);
 598         }
 599         if (src_oa->o_valid & OBD_MD_FLSIZE &&
 600             ret_oa->o_size != src_oa->o_size) {
 601                 CERROR("original size "LPU64" isn't new object size "LPU64"\n",
 602                        src_oa->o_size, ret_oa->o_size);
 603                 LBUG();
 604         }
 605         ret_oa->o_id = src_oa->o_id;
 606         ret_oa->o_gr = src_oa->o_gr;
 607         ret_oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 608         memcpy(src_oa, ret_oa, sizeof(*src_oa));
 609         OBDO_FREE(ret_oa);
 610
 611         *lsmp = set->set_oi->oi_md;
 612         GOTO(done, rc = 0);
 613
 614 cleanup:
 615         list_for_each_entry(req, &set->set_list, rq_link) {
 616                 struct obd_export *sub_exp;
 617                 int err = 0;
 618
 619                 if (!req->rq_complete || req->rq_rc)
 620                         continue;
 621
 622                 sub_exp = lov->lov_tgts[req->rq_idx]->ltd_exp;
 623                 err = obd_destroy(sub_exp, req->rq_oi.oi_oa, NULL, oti, NULL,
 624                                   NULL);
 625                 if (err)
 626                         CERROR("Failed to uncreate objid "LPX64" subobj "
 627                                LPX64" on OST idx %d: rc = %d\n",
 628                                src_oa->o_id, req->rq_oi.oi_oa->o_id,
 629                                req->rq_idx, rc);
 630         }
 631         if (*lsmp == NULL)
 632                 obd_free_memmd(exp, &set->set_oi->oi_md);
 633 done:
 634         if (oti && set->set_cookies) {
 635                 oti->oti_logcookies = set->set_cookies;
 636                 if (!set->set_cookie_sent) {
 637                         oti_free_cookies(oti);
 638                         src_oa->o_valid &= ~OBD_MD_FLCOOKIE;
 639                 } else {
 640                         src_oa->o_valid |= OBD_MD_FLCOOKIE;
 641                 }
 642         }
 643         RETURN(rc);
 644 }
 645
 646 int lov_fini_create_set(struct lov_request_set *set,struct lov_stripe_md **lsmp)
 647 {
 648         int rc = 0;
 649         ENTRY;
 650
 651         if (set == NULL)
 652                 RETURN(0);
 653         LASSERT(set->set_exp);
 654         if (set->set_completes)
 655                 rc = create_done(set->set_exp, set, lsmp);
 656
 657         if (atomic_dec_and_test(&set->set_refcount))
 658                 lov_finish_set(set);
 659
 660         RETURN(rc);
 661 }
 662
 663 int lov_update_create_set(struct lov_request_set *set,
 664                           struct lov_request *req, int rc)
 665 {
 666         struct obd_trans_info *oti = set->set_oti;
 667         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 668         struct lov_oinfo *loi;
 669         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 670         ENTRY;
 671
 672         req->rq_stripe = set->set_success;
 673         loi = lsm->lsm_oinfo[req->rq_stripe];
 674
 675         if (rc && lov->lov_tgts[req->rq_idx] &&
 676             lov->lov_tgts[req->rq_idx]->ltd_active) {
 677                 CERROR("error creating fid "LPX64" sub-object"
 678                        " on OST idx %d/%d: rc = %d\n",
 679                        set->set_oi->oi_oa->o_id, req->rq_idx,
 680                        lsm->lsm_stripe_count, rc);
 681                 if (rc > 0) {
 682                         CERROR("obd_create returned invalid err %d\n", rc);
 683                         rc = -EIO;
 684                 }
 685         }
 686         lov_update_set(set, req, rc);
 687         if (rc)
 688                 RETURN(rc);
 689
 690         loi->loi_id = req->rq_oi.oi_oa->o_id;
 691         loi->loi_gr = req->rq_oi.oi_oa->o_gr;
 692         loi->loi_ost_idx = req->rq_idx;
 693         CDEBUG(D_INODE, "objid "LPX64" has subobj "LPX64"/"LPU64" at idx %d\n",
 694                lsm->lsm_object_id, loi->loi_id, loi->loi_id, req->rq_idx);
 695         loi_init(loi);
 696
 697         if (oti && set->set_cookies)
 698                 ++oti->oti_logcookies;
 699         if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLCOOKIE)
 700                 set->set_cookie_sent++;
 701
 702         RETURN(0);
 703 }
 704
 705 int lov_prep_create_set(struct obd_export *exp, struct obd_info *oinfo,
 706                         struct lov_stripe_md **lsmp, struct obdo *src_oa,
 707                         struct obd_trans_info *oti,
 708                         struct lov_request_set **reqset)
 709 {
 710         struct lov_request_set *set;
 711         int rc = 0;
 712         ENTRY;
 713
 714         OBD_ALLOC(set, sizeof(*set));
 715         if (set == NULL)
 716                 RETURN(-ENOMEM);
 717         lov_init_set(set);
 718
 719         set->set_exp = exp;
 720         set->set_oi = oinfo;
 721         set->set_oi->oi_md = *lsmp;
 722         set->set_oi->oi_oa = src_oa;
 723         set->set_oti = oti;
 724
 725         rc = qos_prep_create(exp, set);
 726         if (rc)
 727                 lov_fini_create_set(set, lsmp);
 728         else
 729                 *reqset = set;
 730         RETURN(rc);
 731 }
 732
 733 static int common_attr_done(struct lov_request_set *set)
 734 {
 735         struct list_head *pos;
 736         struct lov_request *req;
 737         struct obdo *tmp_oa;
 738         int rc = 0, attrset = 0;
 739         ENTRY;
 740
 741         LASSERT(set->set_oi != NULL);
 742
 743         if (set->set_oi->oi_oa == NULL)
 744                 RETURN(0);
 745
 746         if (!set->set_success)
 747                 RETURN(-EIO);
 748
 749         OBDO_ALLOC(tmp_oa);
 750         if (tmp_oa == NULL)
 751                 GOTO(out, rc = -ENOMEM);
 752
 753         list_for_each (pos, &set->set_list) {
 754                 req = list_entry(pos, struct lov_request, rq_link);
 755
 756                 if (!req->rq_complete || req->rq_rc)
 757                         continue;
 758                 if (req->rq_oi.oi_oa->o_valid == 0)   /* inactive stripe */
 759                         continue;
 760                 lov_merge_attrs(tmp_oa, req->rq_oi.oi_oa,
 761                                 req->rq_oi.oi_oa->o_valid,
 762                                 set->set_oi->oi_md, req->rq_stripe, &attrset);
 763         }
 764         if (!attrset) {
 765                 CERROR("No stripes had valid attrs\n");
 766                 rc = -EIO;
 767         }
 768         tmp_oa->o_id = set->set_oi->oi_oa->o_id;
 769         memcpy(set->set_oi->oi_oa, tmp_oa, sizeof(*set->set_oi->oi_oa));
 770 out:
 771         if (tmp_oa)
 772                 OBDO_FREE(tmp_oa);
 773         RETURN(rc);
 774
 775 }
 776
 777 static int brw_done(struct lov_request_set *set)
 778 {
 779         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 780         struct lov_oinfo     *loi = NULL;
 781         struct list_head *pos;
 782         struct lov_request *req;
 783         ENTRY;
 784
 785         list_for_each (pos, &set->set_list) {
 786                 req = list_entry(pos, struct lov_request, rq_link);
 787
 788                 if (!req->rq_complete || req->rq_rc)
 789                         continue;
 790
 791                 loi = lsm->lsm_oinfo[req->rq_stripe];
 792
 793                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLBLOCKS)
 794                         loi->loi_lvb.lvb_blocks = req->rq_oi.oi_oa->o_blocks;
 795         }
 796
 797         RETURN(0);
 798 }
 799
 800 int lov_fini_brw_set(struct lov_request_set *set)
 801 {
 802         int rc = 0;
 803         ENTRY;
 804
 805         if (set == NULL)
 806                 RETURN(0);
 807         LASSERT(set->set_exp);
 808         if (set->set_completes) {
 809                 rc = brw_done(set);
 810                 /* FIXME update qos data here */
 811         }
 812         if (atomic_dec_and_test(&set->set_refcount))
 813                 lov_finish_set(set);
 814
 815         RETURN(rc);
 816 }
 817
 818 int lov_prep_brw_set(struct obd_export *exp, struct obd_info *oinfo,
 819                      obd_count oa_bufs, struct brw_page *pga,
 820                      struct obd_trans_info *oti,
 821                      struct lov_request_set **reqset)
 822 {
 823         struct {
 824                 obd_count       index;
 825                 obd_count       count;
 826                 obd_count       off;
 827         } *info = NULL;
 828         struct lov_request_set *set;
 829         struct lov_obd *lov = &exp->exp_obd->u.lov;
 830         int rc = 0, i, shift;
 831         ENTRY;
 832
 833         OBD_ALLOC(set, sizeof(*set));
 834         if (set == NULL)
 835                 RETURN(-ENOMEM);
 836         lov_init_set(set);
 837
 838         set->set_exp = exp;
 839         set->set_oti = oti;
 840         set->set_oi = oinfo;
 841         set->set_oabufs = oa_bufs;
 842         OBD_ALLOC(set->set_pga, oa_bufs * sizeof(*set->set_pga));
 843         if (!set->set_pga)
 844                 GOTO(out, rc = -ENOMEM);
 845
 846         OBD_ALLOC(info, sizeof(*info) * oinfo->oi_md->lsm_stripe_count);
 847         if (!info)
 848                 GOTO(out, rc = -ENOMEM);
 849
 850         /* calculate the page count for each stripe */
 851         for (i = 0; i < oa_bufs; i++) {
 852                 int stripe = lov_stripe_number(oinfo->oi_md, pga[i].off);
 853                 info[stripe].count++;
 854         }
 855
 856         /* alloc and initialize lov request */
 857         shift = 0;
 858         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++){
 859                 struct lov_oinfo *loi = NULL;
 860                 struct lov_request *req;
 861
 862                 if (info[i].count == 0)
 863                         continue;
 864
 865                 loi = oinfo->oi_md->lsm_oinfo[i];
 866                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 867                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 868                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 869                         GOTO(out, rc = -EIO);
 870                 }
 871
 872                 OBD_ALLOC(req, sizeof(*req));
 873                 if (req == NULL)
 874                         GOTO(out, rc = -ENOMEM);
 875
 876                 OBDO_ALLOC(req->rq_oi.oi_oa);
 877                 if (req->rq_oi.oi_oa == NULL) {
 878                         OBD_FREE(req, sizeof(*req));
 879                         GOTO(out, rc = -ENOMEM);
 880                 }
 881
 882                 if (oinfo->oi_oa) {
 883                         memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
 884                                sizeof(*req->rq_oi.oi_oa));
 885                 }
 886                 req->rq_oi.oi_oa->o_id = loi->loi_id;
 887                 req->rq_oi.oi_oa->o_stripe_idx = i;
 888
 889                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 890                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
 891                 if (req->rq_oi.oi_md == NULL) {
 892                         OBDO_FREE(req->rq_oi.oi_oa);
 893                         OBD_FREE(req, sizeof(*req));
 894                         GOTO(out, rc = -ENOMEM);
 895                 }
 896
 897                 req->rq_idx = loi->loi_ost_idx;
 898                 req->rq_stripe = i;
 899
 900                 /* XXX LOV STACKING */
 901                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 902                 req->rq_oi.oi_md->lsm_object_gr = oinfo->oi_md->lsm_object_gr;
 903                 req->rq_oabufs = info[i].count;
 904                 req->rq_pgaidx = shift;
 905                 shift += req->rq_oabufs;
 906
 907                 /* remember the index for sort brw_page array */
 908                 info[i].index = req->rq_pgaidx;
 909
 910                 req->rq_oi.oi_capa = oinfo->oi_capa;
 911
 912                 lov_set_add_req(req, set);
 913         }
 914         if (!set->set_count)
 915                 GOTO(out, rc = -EIO);
 916
 917         /* rotate & sort the brw_page array */
 918         for (i = 0; i < oa_bufs; i++) {
 919                 int stripe = lov_stripe_number(oinfo->oi_md, pga[i].off);
 920
 921                 shift = info[stripe].index + info[stripe].off;
 922                 LASSERT(shift < oa_bufs);
 923                 set->set_pga[shift] = pga[i];
 924                 lov_stripe_offset(oinfo->oi_md, pga[i].off, stripe,
 925                                   &set->set_pga[shift].off);
 926                 info[stripe].off++;
 927         }
 928 out:
 929         if (info)
 930                 OBD_FREE(info, sizeof(*info) * oinfo->oi_md->lsm_stripe_count);
 931
 932         if (rc == 0)
 933                 *reqset = set;
 934         else
 935                 lov_fini_brw_set(set);
 936
 937         RETURN(rc);
 938 }
 939
 940 int lov_fini_getattr_set(struct lov_request_set *set)
 941 {
 942         int rc = 0;
 943         ENTRY;
 944
 945         if (set == NULL)
 946                 RETURN(0);
 947         LASSERT(set->set_exp);
 948         if (set->set_completes)
 949                 rc = common_attr_done(set);
 950
 951         if (atomic_dec_and_test(&set->set_refcount))
 952                 lov_finish_set(set);
 953
 954         RETURN(rc);
 955 }
 956
 957 /* The callback for osc_getattr_async that finilizes a request info when a
 958  * response is received. */
 959 static int cb_getattr_update(void *cookie, int rc)
 960 {
 961         struct obd_info *oinfo = cookie;
 962         struct lov_request *lovreq;
 963         lovreq = container_of(oinfo, struct lov_request, rq_oi);
 964         return lov_update_common_set(lovreq->rq_rqset, lovreq, rc);
 965 }
 966
 967 int lov_prep_getattr_set(struct obd_export *exp, struct obd_info *oinfo,
 968                          struct lov_request_set **reqset)
 969 {
 970         struct lov_request_set *set;
 971         struct lov_obd *lov = &exp->exp_obd->u.lov;
 972         int rc = 0, i;
 973         ENTRY;
 974
 975         OBD_ALLOC(set, sizeof(*set));
 976         if (set == NULL)
 977                 RETURN(-ENOMEM);
 978         lov_init_set(set);
 979
 980         set->set_exp = exp;
 981         set->set_oi = oinfo;
 982
 983         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
 984                 struct lov_oinfo *loi;
 985                 struct lov_request *req;
 986
 987                 loi = oinfo->oi_md->lsm_oinfo[i];
 988                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 989                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 990                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 991                         continue;
 992                 }
 993
 994                 OBD_ALLOC(req, sizeof(*req));
 995                 if (req == NULL)
 996                         GOTO(out_set, rc = -ENOMEM);
 997
 998                 req->rq_stripe = i;
 999                 req->rq_idx = loi->loi_ost_idx;
1000
1001                 OBDO_ALLOC(req->rq_oi.oi_oa);
1002                 if (req->rq_oi.oi_oa == NULL) {
1003                         OBD_FREE(req, sizeof(*req));
1004                         GOTO(out_set, rc = -ENOMEM);
1005                 }
1006                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1007                        sizeof(*req->rq_oi.oi_oa));
1008                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1009                 req->rq_oi.oi_cb_up = cb_getattr_update;
1010                 req->rq_oi.oi_capa = oinfo->oi_capa;
1011                 req->rq_rqset = set;
1012
1013                 lov_set_add_req(req, set);
1014         }
1015         if (!set->set_count)
1016                 GOTO(out_set, rc = -EIO);
1017         *reqset = set;
1018         RETURN(rc);
1019 out_set:
1020         lov_fini_getattr_set(set);
1021         RETURN(rc);
1022 }
1023
1024 int lov_fini_destroy_set(struct lov_request_set *set)
1025 {
1026         ENTRY;
1027
1028         if (set == NULL)
1029                 RETURN(0);
1030         LASSERT(set->set_exp);
1031         if (set->set_completes) {
1032                 /* FIXME update qos data here */
1033         }
1034
1035         if (atomic_dec_and_test(&set->set_refcount))
1036                 lov_finish_set(set);
1037
1038         RETURN(0);
1039 }
1040
1041 int lov_prep_destroy_set(struct obd_export *exp, struct obd_info *oinfo,
1042                          struct obdo *src_oa, struct lov_stripe_md *lsm,
1043                          struct obd_trans_info *oti,
1044                          struct lov_request_set **reqset)
1045 {
1046         struct lov_request_set *set;
1047         struct lov_obd *lov = &exp->exp_obd->u.lov;
1048         int rc = 0, i;
1049         ENTRY;
1050
1051         OBD_ALLOC(set, sizeof(*set));
1052         if (set == NULL)
1053                 RETURN(-ENOMEM);
1054         lov_init_set(set);
1055
1056         set->set_exp = exp;
1057         set->set_oi = oinfo;
1058         set->set_oi->oi_md = lsm;
1059         set->set_oi->oi_oa = src_oa;
1060         set->set_oti = oti;
1061         if (oti != NULL && src_oa->o_valid & OBD_MD_FLCOOKIE)
1062                 set->set_cookies = oti->oti_logcookies;
1063
1064         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1065                 struct lov_oinfo *loi;
1066                 struct lov_request *req;
1067
1068                 loi = lsm->lsm_oinfo[i];
1069                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1070                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1071                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1072                         continue;
1073                 }
1074
1075                 OBD_ALLOC(req, sizeof(*req));
1076                 if (req == NULL)
1077                         GOTO(out_set, rc = -ENOMEM);
1078
1079                 req->rq_stripe = i;
1080                 req->rq_idx = loi->loi_ost_idx;
1081
1082                 OBDO_ALLOC(req->rq_oi.oi_oa);
1083                 if (req->rq_oi.oi_oa == NULL) {
1084                         OBD_FREE(req, sizeof(*req));
1085                         GOTO(out_set, rc = -ENOMEM);
1086                 }
1087                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1088                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1089                 lov_set_add_req(req, set);
1090         }
1091         if (!set->set_count)
1092                 GOTO(out_set, rc = -EIO);
1093         *reqset = set;
1094         RETURN(rc);
1095 out_set:
1096         lov_fini_destroy_set(set);
1097         RETURN(rc);
1098 }
1099
1100 int lov_fini_setattr_set(struct lov_request_set *set)
1101 {
1102         int rc = 0;
1103         ENTRY;
1104
1105         if (set == NULL)
1106                 RETURN(0);
1107         LASSERT(set->set_exp);
1108         if (set->set_completes) {
1109                 rc = common_attr_done(set);
1110                 /* FIXME update qos data here */
1111         }
1112
1113         if (atomic_dec_and_test(&set->set_refcount))
1114                 lov_finish_set(set);
1115         RETURN(rc);
1116 }
1117
1118 int lov_update_setattr_set(struct lov_request_set *set,
1119                            struct lov_request *req, int rc)
1120 {
1121         struct lov_obd *lov = &req->rq_rqset->set_exp->exp_obd->u.lov;
1122         struct lov_stripe_md *lsm = req->rq_rqset->set_oi->oi_md;
1123         ENTRY;
1124
1125         lov_update_set(set, req, rc);
1126
1127         /* grace error on inactive ost */
1128         if (rc && !(lov->lov_tgts[req->rq_idx] &&
1129                     lov->lov_tgts[req->rq_idx]->ltd_active))
1130                 rc = 0;
1131
1132         if (rc == 0) {
1133                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLCTIME)
1134                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_ctime =
1135                                 req->rq_oi.oi_oa->o_ctime;
1136                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLMTIME)
1137                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_mtime =
1138                                 req->rq_oi.oi_oa->o_mtime;
1139                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLATIME)
1140                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_atime =
1141                                 req->rq_oi.oi_oa->o_atime;
1142         }
1143
1144         RETURN(rc);
1145 }
1146
1147 /* The callback for osc_setattr_async that finilizes a request info when a
1148  * response is received. */
1149 static int cb_setattr_update(void *cookie, int rc)
1150 {
1151         struct obd_info *oinfo = cookie;
1152         struct lov_request *lovreq;
1153         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1154         return lov_update_setattr_set(lovreq->rq_rqset, lovreq, rc);
1155 }
1156
1157 int lov_prep_setattr_set(struct obd_export *exp, struct obd_info *oinfo,
1158                          struct obd_trans_info *oti,
1159                          struct lov_request_set **reqset)
1160 {
1161         struct lov_request_set *set;
1162         struct lov_obd *lov = &exp->exp_obd->u.lov;
1163         int rc = 0, i;
1164         ENTRY;
1165
1166         OBD_ALLOC(set, sizeof(*set));
1167         if (set == NULL)
1168                 RETURN(-ENOMEM);
1169         lov_init_set(set);
1170
1171         set->set_exp = exp;
1172         set->set_oti = oti;
1173         set->set_oi = oinfo;
1174         if (oti != NULL && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
1175                 set->set_cookies = oti->oti_logcookies;
1176
1177         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1178                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1179                 struct lov_request *req;
1180
1181                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1182                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1183                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1184                         continue;
1185                 }
1186
1187                 OBD_ALLOC(req, sizeof(*req));
1188                 if (req == NULL)
1189                         GOTO(out_set, rc = -ENOMEM);
1190                 req->rq_stripe = i;
1191                 req->rq_idx = loi->loi_ost_idx;
1192
1193                 OBDO_ALLOC(req->rq_oi.oi_oa);
1194                 if (req->rq_oi.oi_oa == NULL) {
1195                         OBD_FREE(req, sizeof(*req));
1196                         GOTO(out_set, rc = -ENOMEM);
1197                 }
1198                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1199                        sizeof(*req->rq_oi.oi_oa));
1200                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1201                 LASSERTF(!(req->rq_oi.oi_oa->o_valid & OBD_MD_FLGROUP) ||
1202                          CHECK_MDS_GROUP(req->rq_oi.oi_oa->o_gr),
1203                          "req->rq_oi.oi_oa->o_valid="LPX64" "
1204                          "req->rq_oi.oi_oa->o_gr="LPU64"\n",
1205                          req->rq_oi.oi_oa->o_valid, req->rq_oi.oi_oa->o_gr);
1206                 req->rq_oi.oi_oa->o_stripe_idx = i;
1207                 req->rq_oi.oi_cb_up = cb_setattr_update;
1208                 req->rq_oi.oi_capa = oinfo->oi_capa;
1209                 req->rq_rqset = set;
1210
1211                 if (oinfo->oi_oa->o_valid & OBD_MD_FLSIZE) {
1212                         int off = lov_stripe_offset(oinfo->oi_md,
1213                                                     oinfo->oi_oa->o_size, i,
1214                                                     &req->rq_oi.oi_oa->o_size);
1215
1216                         if (off < 0 && req->rq_oi.oi_oa->o_size)
1217                                 req->rq_oi.oi_oa->o_size--;
1218
1219                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1220                                i, req->rq_oi.oi_oa->o_size,
1221                                oinfo->oi_oa->o_size);
1222                 }
1223                 lov_set_add_req(req, set);
1224         }
1225         if (!set->set_count)
1226                 GOTO(out_set, rc = -EIO);
1227         *reqset = set;
1228         RETURN(rc);
1229 out_set:
1230         lov_fini_setattr_set(set);
1231         RETURN(rc);
1232 }
1233
1234 int lov_fini_punch_set(struct lov_request_set *set)
1235 {
1236         int rc = 0;
1237         ENTRY;
1238
1239         if (set == NULL)
1240                 RETURN(0);
1241         LASSERT(set->set_exp);
1242         if (set->set_completes) {
1243                 rc = -EIO;
1244                 /* FIXME update qos data here */
1245                 if (set->set_success)
1246                         rc = common_attr_done(set);
1247         }
1248
1249         if (atomic_dec_and_test(&set->set_refcount))
1250                 lov_finish_set(set);
1251
1252         RETURN(rc);
1253 }
1254
1255 int lov_update_punch_set(struct lov_request_set *set,
1256                          struct lov_request *req, int rc)
1257 {
1258         struct lov_obd *lov = &req->rq_rqset->set_exp->exp_obd->u.lov;
1259         struct lov_stripe_md *lsm = req->rq_rqset->set_oi->oi_md;
1260         ENTRY;
1261
1262         lov_update_set(set, req, rc);
1263
1264         /* grace error on inactive ost */
1265         if (rc && !lov->lov_tgts[req->rq_idx]->ltd_active)
1266                 rc = 0;
1267
1268         if (rc == 0) {
1269                 lov_stripe_lock(lsm);
1270                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLBLOCKS) {
1271                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_blocks =
1272                                 req->rq_oi.oi_oa->o_blocks;
1273                 }
1274
1275                 /* Do we need to update lvb_size here? It needn't because
1276                  * it have been done in ll_truncate(). -jay */
1277                 lov_stripe_unlock(lsm);
1278         }
1279
1280         RETURN(rc);
1281 }
1282
1283 /* The callback for osc_punch that finilizes a request info when a response
1284  * is received. */
1285 static int cb_update_punch(void *cookie, int rc)
1286 {
1287         struct obd_info *oinfo = cookie;
1288         struct lov_request *lovreq;
1289         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1290         return lov_update_punch_set(lovreq->rq_rqset, lovreq, rc);
1291 }
1292
1293 int lov_prep_punch_set(struct obd_export *exp, struct obd_info *oinfo,
1294                        struct obd_trans_info *oti,
1295                        struct lov_request_set **reqset)
1296 {
1297         struct lov_request_set *set;
1298         struct lov_obd *lov = &exp->exp_obd->u.lov;
1299         int rc = 0, i;
1300         ENTRY;
1301
1302         OBD_ALLOC(set, sizeof(*set));
1303         if (set == NULL)
1304                 RETURN(-ENOMEM);
1305         lov_init_set(set);
1306
1307         set->set_oi = oinfo;
1308         set->set_exp = exp;
1309
1310         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1311                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1312                 struct lov_request *req;
1313                 obd_off rs, re;
1314
1315                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1316                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1317                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1318                         continue;
1319                 }
1320
1321                 if (!lov_stripe_intersects(oinfo->oi_md, i,
1322                                            oinfo->oi_policy.l_extent.start,
1323                                            oinfo->oi_policy.l_extent.end,
1324                                            &rs, &re))
1325                         continue;
1326
1327                 OBD_ALLOC(req, sizeof(*req));
1328                 if (req == NULL)
1329                         GOTO(out_set, rc = -ENOMEM);
1330                 req->rq_stripe = i;
1331                 req->rq_idx = loi->loi_ost_idx;
1332
1333                 OBDO_ALLOC(req->rq_oi.oi_oa);
1334                 if (req->rq_oi.oi_oa == NULL) {
1335                         OBD_FREE(req, sizeof(*req));
1336                         GOTO(out_set, rc = -ENOMEM);
1337                 }
1338                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1339                        sizeof(*req->rq_oi.oi_oa));
1340                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1341                 req->rq_oi.oi_oa->o_gr = loi->loi_gr;
1342                 req->rq_oi.oi_oa->o_valid |= OBD_MD_FLGROUP;
1343
1344                 req->rq_oi.oi_oa->o_stripe_idx = i;
1345                 req->rq_oi.oi_cb_up = cb_update_punch;
1346                 req->rq_rqset = set;
1347
1348                 req->rq_oi.oi_policy.l_extent.start = rs;
1349                 req->rq_oi.oi_policy.l_extent.end = re;
1350                 req->rq_oi.oi_policy.l_extent.gid = -1;
1351
1352                 req->rq_oi.oi_capa = oinfo->oi_capa;
1353
1354                 lov_set_add_req(req, set);
1355         }
1356         if (!set->set_count)
1357                 GOTO(out_set, rc = -EIO);
1358         *reqset = set;
1359         RETURN(rc);
1360 out_set:
1361         lov_fini_punch_set(set);
1362         RETURN(rc);
1363 }
1364
1365 int lov_fini_sync_set(struct lov_request_set *set)
1366 {
1367         int rc = 0;
1368         ENTRY;
1369
1370         if (set == NULL)
1371                 RETURN(0);
1372         LASSERT(set->set_exp);
1373         if (set->set_completes) {
1374                 if (!set->set_success)
1375                         rc = -EIO;
1376                 /* FIXME update qos data here */
1377         }
1378
1379         if (atomic_dec_and_test(&set->set_refcount))
1380                 lov_finish_set(set);
1381
1382         RETURN(rc);
1383 }
1384
1385 int lov_prep_sync_set(struct obd_export *exp, struct obd_info *oinfo,
1386                       struct obdo *src_oa, struct lov_stripe_md *lsm,
1387                       obd_off start, obd_off end,
1388                       struct lov_request_set **reqset)
1389 {
1390         struct lov_request_set *set;
1391         struct lov_obd *lov = &exp->exp_obd->u.lov;
1392         int rc = 0, i;
1393         ENTRY;
1394
1395         OBD_ALLOC(set, sizeof(*set));
1396         if (set == NULL)
1397                 RETURN(-ENOMEM);
1398         lov_init_set(set);
1399
1400         set->set_exp = exp;
1401         set->set_oi = oinfo;
1402         set->set_oi->oi_md = lsm;
1403         set->set_oi->oi_oa = src_oa;
1404
1405         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1406                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
1407                 struct lov_request *req;
1408                 obd_off rs, re;
1409
1410                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1411                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1412                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1413                         continue;
1414                 }
1415
1416                 if (!lov_stripe_intersects(lsm, i, start, end, &rs, &re))
1417                         continue;
1418
1419                 OBD_ALLOC(req, sizeof(*req));
1420                 if (req == NULL)
1421                         GOTO(out_set, rc = -ENOMEM);
1422                 req->rq_stripe = i;
1423                 req->rq_idx = loi->loi_ost_idx;
1424
1425                 OBDO_ALLOC(req->rq_oi.oi_oa);
1426                 if (req->rq_oi.oi_oa == NULL) {
1427                         OBD_FREE(req, sizeof(*req));
1428                         GOTO(out_set, rc = -ENOMEM);
1429                 }
1430                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1431                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1432                 req->rq_oi.oi_oa->o_stripe_idx = i;
1433
1434                 req->rq_oi.oi_policy.l_extent.start = rs;
1435                 req->rq_oi.oi_policy.l_extent.end = re;
1436                 req->rq_oi.oi_policy.l_extent.gid = -1;
1437                 req->rq_rqset = set;
1438
1439                 lov_set_add_req(req, set);
1440         }
1441         if (!set->set_count)
1442                 GOTO(out_set, rc = -EIO);
1443         *reqset = set;
1444         RETURN(rc);
1445 out_set:
1446         lov_fini_sync_set(set);
1447         RETURN(rc);
1448 }
1449
1450 #define LOV_U64_MAX ((__u64)~0ULL)
1451 #define LOV_SUM_MAX(tot, add)                                           \
1452         do {                                                            \
1453                 if ((tot) + (add) < (tot))                              \
1454                         (tot) = LOV_U64_MAX;                            \
1455                 else                                                    \
1456                         (tot) += (add);                                 \
1457         } while(0)
1458
1459 int lov_fini_statfs(struct obd_device *obd, struct obd_statfs *osfs,int success)
1460 {
1461         ENTRY;
1462
1463         if (success) {
1464                 __u32 expected_stripes = lov_get_stripecnt(&obd->u.lov, 0);
1465
1466                 if (osfs->os_files != LOV_U64_MAX)
1467                         do_div(osfs->os_files, expected_stripes);
1468                 if (osfs->os_ffree != LOV_U64_MAX)
1469                         do_div(osfs->os_ffree, expected_stripes);
1470
1471                 spin_lock(&obd->obd_osfs_lock);
1472                 memcpy(&obd->obd_osfs, osfs, sizeof(*osfs));
1473                 obd->obd_osfs_age = cfs_time_current_64();
1474                 spin_unlock(&obd->obd_osfs_lock);
1475                 RETURN(0);
1476         }
1477
1478         RETURN(-EIO);
1479 }
1480
1481 int lov_fini_statfs_set(struct lov_request_set *set)
1482 {
1483         int rc = 0;
1484         ENTRY;
1485
1486         if (set == NULL)
1487                 RETURN(0);
1488
1489         if (set->set_completes) {
1490                 rc = lov_fini_statfs(set->set_obd, set->set_oi->oi_osfs,
1491                                      set->set_success);
1492         }
1493
1494         if (atomic_dec_and_test(&set->set_refcount))
1495                 lov_finish_set(set);
1496
1497         RETURN(rc);
1498 }
1499
1500 void lov_update_statfs(struct obd_statfs *osfs, struct obd_statfs *lov_sfs,
1501                        int success)
1502 {
1503         int shift = 0, quit = 0;
1504         __u64 tmp;
1505
1506         if (success == 0) {
1507                 memcpy(osfs, lov_sfs, sizeof(*lov_sfs));
1508         } else {
1509                 if (osfs->os_bsize != lov_sfs->os_bsize) {
1510                         /* assume all block sizes are always powers of 2 */
1511                         /* get the bits difference */
1512                         tmp = osfs->os_bsize | lov_sfs->os_bsize;
1513                         for (shift = 0; shift <= 64; ++shift) {
1514                                 if (tmp & 1) {
1515                                         if (quit)
1516                                                 break;
1517                                         else
1518                                                 quit = 1;
1519                                         shift = 0;
1520                                 }
1521                                 tmp >>= 1;
1522                         }
1523                 }
1524
1525                 if (osfs->os_bsize < lov_sfs->os_bsize) {
1526                         osfs->os_bsize = lov_sfs->os_bsize;
1527
1528                         osfs->os_bfree  >>= shift;
1529                         osfs->os_bavail >>= shift;
1530                         osfs->os_blocks >>= shift;
1531                 } else if (shift != 0) {
1532                         lov_sfs->os_bfree  >>= shift;
1533                         lov_sfs->os_bavail >>= shift;
1534                         lov_sfs->os_blocks >>= shift;
1535                 }
1536 #ifdef MIN_DF
1537                 /* Sandia requested that df (and so, statfs) only
1538                    returned minimal available space on
1539                    a single OST, so people would be able to
1540                    write this much data guaranteed. */
1541                 if (osfs->os_bavail > lov_sfs->os_bavail) {
1542                         /* Presumably if new bavail is smaller,
1543                            new bfree is bigger as well */
1544                         osfs->os_bfree = lov_sfs->os_bfree;
1545                         osfs->os_bavail = lov_sfs->os_bavail;
1546                 }
1547 #else
1548                 osfs->os_bfree += lov_sfs->os_bfree;
1549                 osfs->os_bavail += lov_sfs->os_bavail;
1550 #endif
1551                 osfs->os_blocks += lov_sfs->os_blocks;
1552                 /* XXX not sure about this one - depends on policy.
1553                  *   - could be minimum if we always stripe on all OBDs
1554                  *     (but that would be wrong for any other policy,
1555                  *     if one of the OBDs has no more objects left)
1556                  *   - could be sum if we stripe whole objects
1557                  *   - could be average, just to give a nice number
1558                  *
1559                  * To give a "reasonable" (if not wholly accurate)
1560                  * number, we divide the total number of free objects
1561                  * by expected stripe count (watch out for overflow).
1562                  */
1563                 LOV_SUM_MAX(osfs->os_files, lov_sfs->os_files);
1564                 LOV_SUM_MAX(osfs->os_ffree, lov_sfs->os_ffree);
1565         }
1566 }
1567
1568 /* The callback for osc_statfs_async that finilizes a request info when a
1569  * response is received. */
1570 static int cb_statfs_update(void *cookie, int rc)
1571 {
1572         struct obd_info *oinfo = cookie;
1573         struct lov_request *lovreq;
1574         struct obd_statfs *osfs, *lov_sfs;
1575         struct obd_device *obd;
1576         struct lov_obd *lov;
1577         int success;
1578         ENTRY;
1579
1580         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1581         lov = &lovreq->rq_rqset->set_obd->u.lov;
1582         obd = class_exp2obd(lov->lov_tgts[lovreq->rq_idx]->ltd_exp);
1583
1584         osfs = lovreq->rq_rqset->set_oi->oi_osfs;
1585         lov_sfs = oinfo->oi_osfs;
1586
1587         success = lovreq->rq_rqset->set_success;
1588
1589         /* XXX: the same is done in lov_update_common_set, however
1590            lovset->set_exp is not initialized. */
1591         lov_update_set(lovreq->rq_rqset, lovreq, rc);
1592         if (rc) {
1593                 if (rc && !(lov->lov_tgts[lovreq->rq_idx] &&
1594                             lov->lov_tgts[lovreq->rq_idx]->ltd_active))
1595                         rc = 0;
1596                 GOTO(out, rc);
1597         }
1598
1599         spin_lock(&obd->obd_osfs_lock);
1600         memcpy(&obd->obd_osfs, lov_sfs, sizeof(*lov_sfs));
1601         if ((oinfo->oi_flags & OBD_STATFS_FROM_CACHE) == 0)
1602                 obd->obd_osfs_age = cfs_time_current_64();
1603         spin_unlock(&obd->obd_osfs_lock);
1604
1605         lov_update_statfs(osfs, lov_sfs, success);
1606         qos_update(lov);
1607 out:
1608         if (lovreq->rq_rqset->set_oi->oi_flags & OBD_STATFS_PTLRPCD &&
1609             lovreq->rq_rqset->set_count == lovreq->rq_rqset->set_completes) {
1610                lov_statfs_interpret(NULL, lovreq->rq_rqset,
1611                                     lovreq->rq_rqset->set_success !=
1612                                                   lovreq->rq_rqset->set_count);
1613                qos_statfs_done(lov);
1614         }
1615
1616         RETURN(0);
1617 }
1618
1619 int lov_prep_statfs_set(struct obd_device *obd, struct obd_info *oinfo,
1620                         struct lov_request_set **reqset)
1621 {
1622         struct lov_request_set *set;
1623         struct lov_obd *lov = &obd->u.lov;
1624         int rc = 0, i;
1625         ENTRY;
1626
1627         OBD_ALLOC(set, sizeof(*set));
1628         if (set == NULL)
1629                 RETURN(-ENOMEM);
1630         lov_init_set(set);
1631
1632         set->set_obd = obd;
1633         set->set_oi = oinfo;
1634
1635         /* We only get block data from the OBD */
1636         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1637                 struct lov_request *req;
1638
1639                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active) {
1640                         CDEBUG(D_HA, "lov idx %d inactive\n", i);
1641                         continue;
1642                 }
1643
1644                 OBD_ALLOC(req, sizeof(*req));
1645                 if (req == NULL)
1646                         GOTO(out_set, rc = -ENOMEM);
1647
1648                 OBD_ALLOC(req->rq_oi.oi_osfs, sizeof(*req->rq_oi.oi_osfs));
1649                 if (req->rq_oi.oi_osfs == NULL) {
1650                         OBD_FREE(req, sizeof(*req));
1651                         GOTO(out_set, rc = -ENOMEM);
1652                 }
1653
1654                 req->rq_idx = i;
1655                 req->rq_oi.oi_cb_up = cb_statfs_update;
1656                 req->rq_oi.oi_flags = oinfo->oi_flags;
1657                 req->rq_rqset = set;
1658
1659                 lov_set_add_req(req, set);
1660         }
1661         if (!set->set_count)
1662                 GOTO(out_set, rc = -EIO);
1663         *reqset = set;
1664         RETURN(rc);
1665 out_set:
1666         lov_fini_statfs_set(set);
1667         RETURN(rc);
1668 }