drivers/staging/lustre/lustre/lov/lov_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  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 #define DEBUG_SUBSYSTEM S_LOV
  38
  39 #include "../../include/linux/libcfs/libcfs.h"
  40
  41 #include "../include/obd_class.h"
  42 #include "../include/obd_ost.h"
  43 #include "../include/lustre/lustre_idl.h"
  44 #include "lov_internal.h"
  45
  46 static void lov_init_set(struct lov_request_set *set)
  47 {
  48         set->set_count = 0;
  49         atomic_set(&set->set_completes, 0);
  50         atomic_set(&set->set_success, 0);
  51         atomic_set(&set->set_finish_checked, 0);
  52         set->set_cookies = NULL;
  53         INIT_LIST_HEAD(&set->set_list);
  54         atomic_set(&set->set_refcount, 1);
  55         init_waitqueue_head(&set->set_waitq);
  56         spin_lock_init(&set->set_lock);
  57 }
  58
  59 void lov_finish_set(struct lov_request_set *set)
  60 {
  61         struct list_head *pos, *n;
  62
  63         LASSERT(set);
  64         list_for_each_safe(pos, n, &set->set_list) {
  65                 struct lov_request *req = list_entry(pos,
  66                                                          struct lov_request,
  67                                                          rq_link);
  68                 list_del_init(&req->rq_link);
  69
  70                 if (req->rq_oi.oi_oa)
  71                         OBDO_FREE(req->rq_oi.oi_oa);
  72                 if (req->rq_oi.oi_md)
  73                         OBD_FREE_LARGE(req->rq_oi.oi_md, req->rq_buflen);
  74                 if (req->rq_oi.oi_osfs)
  75                         OBD_FREE(req->rq_oi.oi_osfs,
  76                                  sizeof(*req->rq_oi.oi_osfs));
  77                 OBD_FREE(req, sizeof(*req));
  78         }
  79
  80         if (set->set_pga) {
  81                 int len = set->set_oabufs * sizeof(*set->set_pga);
  82                 OBD_FREE_LARGE(set->set_pga, len);
  83         }
  84         if (set->set_lockh)
  85                 lov_llh_put(set->set_lockh);
  86
  87         OBD_FREE(set, sizeof(*set));
  88 }
  89
  90 int lov_set_finished(struct lov_request_set *set, int idempotent)
  91 {
  92         int completes = atomic_read(&set->set_completes);
  93
  94         CDEBUG(D_INFO, "check set %d/%d\n", completes, set->set_count);
  95
  96         if (completes == set->set_count) {
  97                 if (idempotent)
  98                         return 1;
  99                 if (atomic_inc_return(&set->set_finish_checked) == 1)
 100                         return 1;
 101         }
 102         return 0;
 103 }
 104
 105 void lov_update_set(struct lov_request_set *set,
 106                     struct lov_request *req, int rc)
 107 {
 108         req->rq_complete = 1;
 109         req->rq_rc = rc;
 110
 111         atomic_inc(&set->set_completes);
 112         if (rc == 0)
 113                 atomic_inc(&set->set_success);
 114
 115         wake_up(&set->set_waitq);
 116 }
 117
 118 int lov_update_common_set(struct lov_request_set *set,
 119                           struct lov_request *req, int rc)
 120 {
 121         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 122
 123         lov_update_set(set, req, rc);
 124
 125         /* grace error on inactive ost */
 126         if (rc && !(lov->lov_tgts[req->rq_idx] &&
 127                     lov->lov_tgts[req->rq_idx]->ltd_active))
 128                 rc = 0;
 129
 130         /* FIXME in raid1 regime, should return 0 */
 131         return rc;
 132 }
 133
 134 void lov_set_add_req(struct lov_request *req, struct lov_request_set *set)
 135 {
 136         list_add_tail(&req->rq_link, &set->set_list);
 137         set->set_count++;
 138         req->rq_rqset = set;
 139 }
 140
 141 static int lov_check_set(struct lov_obd *lov, int idx)
 142 {
 143         int rc;
 144         struct lov_tgt_desc *tgt;
 145
 146         mutex_lock(&lov->lov_lock);
 147         tgt = lov->lov_tgts[idx];
 148         rc = !tgt || tgt->ltd_active ||
 149                 (tgt->ltd_exp &&
 150                  class_exp2cliimp(tgt->ltd_exp)->imp_connect_tried);
 151         mutex_unlock(&lov->lov_lock);
 152
 153         return rc;
 154 }
 155
 156 /* Check if the OSC connection exists and is active.
 157  * If the OSC has not yet had a chance to connect to the OST the first time,
 158  * wait once for it to connect instead of returning an error.
 159  */
 160 int lov_check_and_wait_active(struct lov_obd *lov, int ost_idx)
 161 {
 162         wait_queue_head_t waitq;
 163         struct l_wait_info lwi;
 164         struct lov_tgt_desc *tgt;
 165         int rc = 0;
 166
 167         mutex_lock(&lov->lov_lock);
 168
 169         tgt = lov->lov_tgts[ost_idx];
 170
 171         if (unlikely(tgt == NULL))
 172                 GOTO(out, rc = 0);
 173
 174         if (likely(tgt->ltd_active))
 175                 GOTO(out, rc = 1);
 176
 177         if (tgt->ltd_exp && class_exp2cliimp(tgt->ltd_exp)->imp_connect_tried)
 178                 GOTO(out, rc = 0);
 179
 180         mutex_unlock(&lov->lov_lock);
 181
 182         init_waitqueue_head(&waitq);
 183         lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(obd_timeout),
 184                                    cfs_time_seconds(1), NULL, NULL);
 185
 186         rc = l_wait_event(waitq, lov_check_set(lov, ost_idx), &lwi);
 187         if (tgt != NULL && tgt->ltd_active)
 188                 return 1;
 189
 190         return 0;
 191
 192 out:
 193         mutex_unlock(&lov->lov_lock);
 194         return rc;
 195 }
 196
 197 static int lov_update_enqueue_lov(struct obd_export *exp,
 198                                   struct lustre_handle *lov_lockhp,
 199                                   struct lov_oinfo *loi, __u64 flags, int idx,
 200                                   struct ost_id *oi, int rc)
 201 {
 202         struct lov_obd *lov = &exp->exp_obd->u.lov;
 203
 204         if (rc != ELDLM_OK &&
 205             !(rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT))) {
 206                 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
 207                 if (lov->lov_tgts[idx] && lov->lov_tgts[idx]->ltd_active) {
 208                         /* -EUSERS used by OST to report file contention */
 209                         if (rc != -EINTR && rc != -EUSERS)
 210                                 CERROR("%s: enqueue objid "DOSTID" subobj"
 211                                        DOSTID" on OST idx %d: rc %d\n",
 212                                        exp->exp_obd->obd_name,
 213                                        POSTID(oi), POSTID(&loi->loi_oi),
 214                                        loi->loi_ost_idx, rc);
 215                 } else
 216                         rc = ELDLM_OK;
 217         }
 218         return rc;
 219 }
 220
 221 int lov_update_enqueue_set(struct lov_request *req, __u32 mode, int rc)
 222 {
 223         struct lov_request_set *set = req->rq_rqset;
 224         struct lustre_handle *lov_lockhp;
 225         struct obd_info *oi = set->set_oi;
 226         struct lov_oinfo *loi;
 227
 228         LASSERT(oi != NULL);
 229
 230         lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
 231         loi = oi->oi_md->lsm_oinfo[req->rq_stripe];
 232
 233         /* XXX LOV STACKING: OSC gets a copy, created in lov_prep_enqueue_set
 234          * and that copy can be arbitrarily out of date.
 235          *
 236          * The LOV API is due for a serious rewriting anyways, and this
 237          * can be addressed then. */
 238
 239         lov_stripe_lock(oi->oi_md);
 240         osc_update_enqueue(lov_lockhp, loi, oi->oi_flags,
 241                            &req->rq_oi.oi_md->lsm_oinfo[0]->loi_lvb, mode, rc);
 242         if (rc == ELDLM_LOCK_ABORTED && (oi->oi_flags & LDLM_FL_HAS_INTENT))
 243                 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
 244         rc = lov_update_enqueue_lov(set->set_exp, lov_lockhp, loi, oi->oi_flags,
 245                                     req->rq_idx, &oi->oi_md->lsm_oi, rc);
 246         lov_stripe_unlock(oi->oi_md);
 247         lov_update_set(set, req, rc);
 248         return rc;
 249 }
 250
 251 /* The callback for osc_enqueue that updates lov info for every OSC request. */
 252 static int cb_update_enqueue(void *cookie, int rc)
 253 {
 254         struct obd_info *oinfo = cookie;
 255         struct ldlm_enqueue_info *einfo;
 256         struct lov_request *lovreq;
 257
 258         lovreq = container_of(oinfo, struct lov_request, rq_oi);
 259         einfo = lovreq->rq_rqset->set_ei;
 260         return lov_update_enqueue_set(lovreq, einfo->ei_mode, rc);
 261 }
 262
 263 static int enqueue_done(struct lov_request_set *set, __u32 mode)
 264 {
 265         struct lov_request *req;
 266         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 267         int completes = atomic_read(&set->set_completes);
 268         int rc = 0;
 269
 270         /* enqueue/match success, just return */
 271         if (completes && completes == atomic_read(&set->set_success))
 272                 return 0;
 273
 274         /* cancel enqueued/matched locks */
 275         list_for_each_entry(req, &set->set_list, rq_link) {
 276                 struct lustre_handle *lov_lockhp;
 277
 278                 if (!req->rq_complete || req->rq_rc)
 279                         continue;
 280
 281                 lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
 282                 LASSERT(lov_lockhp);
 283                 if (!lustre_handle_is_used(lov_lockhp))
 284                         continue;
 285
 286                 rc = obd_cancel(lov->lov_tgts[req->rq_idx]->ltd_exp,
 287                                 req->rq_oi.oi_md, mode, lov_lockhp);
 288                 if (rc && lov->lov_tgts[req->rq_idx] &&
 289                     lov->lov_tgts[req->rq_idx]->ltd_active)
 290                         CERROR("%s: cancelling obdjid "DOSTID" on OST"
 291                                "idx %d error: rc = %d\n",
 292                                set->set_exp->exp_obd->obd_name,
 293                                POSTID(&req->rq_oi.oi_md->lsm_oi),
 294                                req->rq_idx, rc);
 295         }
 296         if (set->set_lockh)
 297                 lov_llh_put(set->set_lockh);
 298         return rc;
 299 }
 300
 301 int lov_fini_enqueue_set(struct lov_request_set *set, __u32 mode, int rc,
 302                          struct ptlrpc_request_set *rqset)
 303 {
 304         int ret = 0;
 305
 306         if (set == NULL)
 307                 return 0;
 308         LASSERT(set->set_exp);
 309         /* Do enqueue_done only for sync requests and if any request
 310          * succeeded. */
 311         if (!rqset) {
 312                 if (rc)
 313                         atomic_set(&set->set_completes, 0);
 314                 ret = enqueue_done(set, mode);
 315         } else if (set->set_lockh)
 316                 lov_llh_put(set->set_lockh);
 317
 318         lov_put_reqset(set);
 319
 320         return rc ? rc : ret;
 321 }
 322
 323 static void lov_llh_addref(void *llhp)
 324 {
 325         struct lov_lock_handles *llh = llhp;
 326
 327         atomic_inc(&llh->llh_refcount);
 328         CDEBUG(D_INFO, "GETting llh %p : new refcount %d\n", llh,
 329                atomic_read(&llh->llh_refcount));
 330 }
 331
 332 static struct portals_handle_ops lov_handle_ops = {
 333         .hop_addref = lov_llh_addref,
 334         .hop_free   = NULL,
 335 };
 336
 337 static struct lov_lock_handles *lov_llh_new(struct lov_stripe_md *lsm)
 338 {
 339         struct lov_lock_handles *llh;
 340
 341         OBD_ALLOC(llh, sizeof(*llh) +
 342                   sizeof(*llh->llh_handles) * lsm->lsm_stripe_count);
 343         if (llh == NULL)
 344                 return NULL;
 345
 346         atomic_set(&llh->llh_refcount, 2);
 347         llh->llh_stripe_count = lsm->lsm_stripe_count;
 348         INIT_LIST_HEAD(&llh->llh_handle.h_link);
 349         class_handle_hash(&llh->llh_handle, &lov_handle_ops);
 350
 351         return llh;
 352 }
 353
 354 int lov_prep_enqueue_set(struct obd_export *exp, struct obd_info *oinfo,
 355                          struct ldlm_enqueue_info *einfo,
 356                          struct lov_request_set **reqset)
 357 {
 358         struct lov_obd *lov = &exp->exp_obd->u.lov;
 359         struct lov_request_set *set;
 360         int i, rc = 0;
 361
 362         OBD_ALLOC(set, sizeof(*set));
 363         if (set == NULL)
 364                 return -ENOMEM;
 365         lov_init_set(set);
 366
 367         set->set_exp = exp;
 368         set->set_oi = oinfo;
 369         set->set_ei = einfo;
 370         set->set_lockh = lov_llh_new(oinfo->oi_md);
 371         if (set->set_lockh == NULL)
 372                 GOTO(out_set, rc = -ENOMEM);
 373         oinfo->oi_lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 374
 375         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
 376                 struct lov_oinfo *loi;
 377                 struct lov_request *req;
 378                 obd_off start, end;
 379
 380                 loi = oinfo->oi_md->lsm_oinfo[i];
 381                 if (!lov_stripe_intersects(oinfo->oi_md, i,
 382                                            oinfo->oi_policy.l_extent.start,
 383                                            oinfo->oi_policy.l_extent.end,
 384                                            &start, &end))
 385                         continue;
 386
 387                 if (!lov_check_and_wait_active(lov, loi->loi_ost_idx)) {
 388                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 389                         continue;
 390                 }
 391
 392                 OBD_ALLOC(req, sizeof(*req));
 393                 if (req == NULL)
 394                         GOTO(out_set, rc = -ENOMEM);
 395
 396                 req->rq_buflen = sizeof(*req->rq_oi.oi_md) +
 397                         sizeof(struct lov_oinfo *) +
 398                         sizeof(struct lov_oinfo);
 399                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 400                 if (req->rq_oi.oi_md == NULL) {
 401                         OBD_FREE(req, sizeof(*req));
 402                         GOTO(out_set, rc = -ENOMEM);
 403                 }
 404                 req->rq_oi.oi_md->lsm_oinfo[0] =
 405                         ((void *)req->rq_oi.oi_md) + sizeof(*req->rq_oi.oi_md) +
 406                         sizeof(struct lov_oinfo *);
 407
 408                 /* Set lov request specific parameters. */
 409                 req->rq_oi.oi_lockh = set->set_lockh->llh_handles + i;
 410                 req->rq_oi.oi_cb_up = cb_update_enqueue;
 411                 req->rq_oi.oi_flags = oinfo->oi_flags;
 412
 413                 LASSERT(req->rq_oi.oi_lockh);
 414
 415                 req->rq_oi.oi_policy.l_extent.gid =
 416                         oinfo->oi_policy.l_extent.gid;
 417                 req->rq_oi.oi_policy.l_extent.start = start;
 418                 req->rq_oi.oi_policy.l_extent.end = end;
 419
 420                 req->rq_idx = loi->loi_ost_idx;
 421                 req->rq_stripe = i;
 422
 423                 /* XXX LOV STACKING: submd should be from the subobj */
 424                 req->rq_oi.oi_md->lsm_oi = loi->loi_oi;
 425                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 426                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_kms_valid =
 427                         loi->loi_kms_valid;
 428                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_kms = loi->loi_kms;
 429                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_lvb = loi->loi_lvb;
 430
 431                 lov_set_add_req(req, set);
 432         }
 433         if (!set->set_count)
 434                 GOTO(out_set, rc = -EIO);
 435         *reqset = set;
 436         return 0;
 437 out_set:
 438         lov_fini_enqueue_set(set, einfo->ei_mode, rc, NULL);
 439         return rc;
 440 }
 441
 442 int lov_fini_match_set(struct lov_request_set *set, __u32 mode, __u64 flags)
 443 {
 444         int rc = 0;
 445
 446         if (set == NULL)
 447                 return 0;
 448         LASSERT(set->set_exp);
 449         rc = enqueue_done(set, mode);
 450         if ((set->set_count == atomic_read(&set->set_success)) &&
 451             (flags & LDLM_FL_TEST_LOCK))
 452                 lov_llh_put(set->set_lockh);
 453
 454         lov_put_reqset(set);
 455
 456         return rc;
 457 }
 458
 459 int lov_prep_match_set(struct obd_export *exp, struct obd_info *oinfo,
 460                        struct lov_stripe_md *lsm, ldlm_policy_data_t *policy,
 461                        __u32 mode, struct lustre_handle *lockh,
 462                        struct lov_request_set **reqset)
 463 {
 464         struct lov_obd *lov = &exp->exp_obd->u.lov;
 465         struct lov_request_set *set;
 466         int i, rc = 0;
 467
 468         OBD_ALLOC(set, sizeof(*set));
 469         if (set == NULL)
 470                 return -ENOMEM;
 471         lov_init_set(set);
 472
 473         set->set_exp = exp;
 474         set->set_oi = oinfo;
 475         set->set_oi->oi_md = lsm;
 476         set->set_lockh = lov_llh_new(lsm);
 477         if (set->set_lockh == NULL)
 478                 GOTO(out_set, rc = -ENOMEM);
 479         lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 480
 481         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 482                 struct lov_oinfo *loi;
 483                 struct lov_request *req;
 484                 obd_off start, end;
 485
 486                 loi = lsm->lsm_oinfo[i];
 487                 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
 488                                            policy->l_extent.end, &start, &end))
 489                         continue;
 490
 491                 /* FIXME raid1 should grace this error */
 492                 if (!lov_check_and_wait_active(lov, loi->loi_ost_idx)) {
 493                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 494                         GOTO(out_set, rc = -EIO);
 495                 }
 496
 497                 OBD_ALLOC(req, sizeof(*req));
 498                 if (req == NULL)
 499                         GOTO(out_set, rc = -ENOMEM);
 500
 501                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 502                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 503                 if (req->rq_oi.oi_md == NULL) {
 504                         OBD_FREE(req, sizeof(*req));
 505                         GOTO(out_set, rc = -ENOMEM);
 506                 }
 507
 508                 req->rq_oi.oi_policy.l_extent.start = start;
 509                 req->rq_oi.oi_policy.l_extent.end = end;
 510                 req->rq_oi.oi_policy.l_extent.gid = policy->l_extent.gid;
 511
 512                 req->rq_idx = loi->loi_ost_idx;
 513                 req->rq_stripe = i;
 514
 515                 /* XXX LOV STACKING: submd should be from the subobj */
 516                 req->rq_oi.oi_md->lsm_oi = loi->loi_oi;
 517                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 518
 519                 lov_set_add_req(req, set);
 520         }
 521         if (!set->set_count)
 522                 GOTO(out_set, rc = -EIO);
 523         *reqset = set;
 524         return rc;
 525 out_set:
 526         lov_fini_match_set(set, mode, 0);
 527         return rc;
 528 }
 529
 530 int lov_fini_cancel_set(struct lov_request_set *set)
 531 {
 532         int rc = 0;
 533
 534         if (set == NULL)
 535                 return 0;
 536
 537         LASSERT(set->set_exp);
 538         if (set->set_lockh)
 539                 lov_llh_put(set->set_lockh);
 540
 541         lov_put_reqset(set);
 542
 543         return rc;
 544 }
 545
 546 int lov_prep_cancel_set(struct obd_export *exp, struct obd_info *oinfo,
 547                         struct lov_stripe_md *lsm, __u32 mode,
 548                         struct lustre_handle *lockh,
 549                         struct lov_request_set **reqset)
 550 {
 551         struct lov_request_set *set;
 552         int i, rc = 0;
 553
 554         OBD_ALLOC(set, sizeof(*set));
 555         if (set == NULL)
 556                 return -ENOMEM;
 557         lov_init_set(set);
 558
 559         set->set_exp = exp;
 560         set->set_oi = oinfo;
 561         set->set_oi->oi_md = lsm;
 562         set->set_lockh = lov_handle2llh(lockh);
 563         if (set->set_lockh == NULL) {
 564                 CERROR("LOV: invalid lov lock handle %p\n", lockh);
 565                 GOTO(out_set, rc = -EINVAL);
 566         }
 567         lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 568
 569         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 570                 struct lov_request *req;
 571                 struct lustre_handle *lov_lockhp;
 572                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 573
 574                 lov_lockhp = set->set_lockh->llh_handles + i;
 575                 if (!lustre_handle_is_used(lov_lockhp)) {
 576                         CDEBUG(D_INFO, "lov idx %d subobj "DOSTID" no lock\n",
 577                                loi->loi_ost_idx, POSTID(&loi->loi_oi));
 578                         continue;
 579                 }
 580
 581                 OBD_ALLOC(req, sizeof(*req));
 582                 if (req == NULL)
 583                         GOTO(out_set, rc = -ENOMEM);
 584
 585                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 586                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 587                 if (req->rq_oi.oi_md == NULL) {
 588                         OBD_FREE(req, sizeof(*req));
 589                         GOTO(out_set, rc = -ENOMEM);
 590                 }
 591
 592                 req->rq_idx = loi->loi_ost_idx;
 593                 req->rq_stripe = i;
 594
 595                 /* XXX LOV STACKING: submd should be from the subobj */
 596                 req->rq_oi.oi_md->lsm_oi = loi->loi_oi;
 597                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 598
 599                 lov_set_add_req(req, set);
 600         }
 601         if (!set->set_count)
 602                 GOTO(out_set, rc = -EIO);
 603         *reqset = set;
 604         return rc;
 605 out_set:
 606         lov_fini_cancel_set(set);
 607         return rc;
 608 }
 609 static int common_attr_done(struct lov_request_set *set)
 610 {
 611         struct list_head *pos;
 612         struct lov_request *req;
 613         struct obdo *tmp_oa;
 614         int rc = 0, attrset = 0;
 615
 616         LASSERT(set->set_oi != NULL);
 617
 618         if (set->set_oi->oi_oa == NULL)
 619                 return 0;
 620
 621         if (!atomic_read(&set->set_success))
 622                 return -EIO;
 623
 624         OBDO_ALLOC(tmp_oa);
 625         if (tmp_oa == NULL)
 626                 GOTO(out, rc = -ENOMEM);
 627
 628         list_for_each(pos, &set->set_list) {
 629                 req = list_entry(pos, struct lov_request, rq_link);
 630
 631                 if (!req->rq_complete || req->rq_rc)
 632                         continue;
 633                 if (req->rq_oi.oi_oa->o_valid == 0)   /* inactive stripe */
 634                         continue;
 635                 lov_merge_attrs(tmp_oa, req->rq_oi.oi_oa,
 636                                 req->rq_oi.oi_oa->o_valid,
 637                                 set->set_oi->oi_md, req->rq_stripe, &attrset);
 638         }
 639         if (!attrset) {
 640                 CERROR("No stripes had valid attrs\n");
 641                 rc = -EIO;
 642         }
 643         if ((set->set_oi->oi_oa->o_valid & OBD_MD_FLEPOCH) &&
 644             (set->set_oi->oi_md->lsm_stripe_count != attrset)) {
 645                 /* When we take attributes of some epoch, we require all the
 646                  * ost to be active. */
 647                 CERROR("Not all the stripes had valid attrs\n");
 648                 GOTO(out, rc = -EIO);
 649         }
 650
 651         tmp_oa->o_oi = set->set_oi->oi_oa->o_oi;
 652         memcpy(set->set_oi->oi_oa, tmp_oa, sizeof(*set->set_oi->oi_oa));
 653 out:
 654         if (tmp_oa)
 655                 OBDO_FREE(tmp_oa);
 656         return rc;
 657
 658 }
 659
 660 static int brw_done(struct lov_request_set *set)
 661 {
 662         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 663         struct lov_oinfo     *loi = NULL;
 664         struct list_head *pos;
 665         struct lov_request *req;
 666
 667         list_for_each(pos, &set->set_list) {
 668                 req = list_entry(pos, struct lov_request, rq_link);
 669
 670                 if (!req->rq_complete || req->rq_rc)
 671                         continue;
 672
 673                 loi = lsm->lsm_oinfo[req->rq_stripe];
 674
 675                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLBLOCKS)
 676                         loi->loi_lvb.lvb_blocks = req->rq_oi.oi_oa->o_blocks;
 677         }
 678
 679         return 0;
 680 }
 681
 682 int lov_fini_brw_set(struct lov_request_set *set)
 683 {
 684         int rc = 0;
 685
 686         if (set == NULL)
 687                 return 0;
 688         LASSERT(set->set_exp);
 689         if (atomic_read(&set->set_completes)) {
 690                 rc = brw_done(set);
 691                 /* FIXME update qos data here */
 692         }
 693         lov_put_reqset(set);
 694
 695         return rc;
 696 }
 697
 698 int lov_prep_brw_set(struct obd_export *exp, struct obd_info *oinfo,
 699                      obd_count oa_bufs, struct brw_page *pga,
 700                      struct obd_trans_info *oti,
 701                      struct lov_request_set **reqset)
 702 {
 703         struct {
 704                 obd_count       index;
 705                 obd_count       count;
 706                 obd_count       off;
 707         } *info = NULL;
 708         struct lov_request_set *set;
 709         struct lov_obd *lov = &exp->exp_obd->u.lov;
 710         int rc = 0, i, shift;
 711
 712         OBD_ALLOC(set, sizeof(*set));
 713         if (set == NULL)
 714                 return -ENOMEM;
 715         lov_init_set(set);
 716
 717         set->set_exp = exp;
 718         set->set_oti = oti;
 719         set->set_oi = oinfo;
 720         set->set_oabufs = oa_bufs;
 721         OBD_ALLOC_LARGE(set->set_pga, oa_bufs * sizeof(*set->set_pga));
 722         if (!set->set_pga)
 723                 GOTO(out, rc = -ENOMEM);
 724
 725         OBD_ALLOC_LARGE(info, sizeof(*info) * oinfo->oi_md->lsm_stripe_count);
 726         if (!info)
 727                 GOTO(out, rc = -ENOMEM);
 728
 729         /* calculate the page count for each stripe */
 730         for (i = 0; i < oa_bufs; i++) {
 731                 int stripe = lov_stripe_number(oinfo->oi_md, pga[i].off);
 732                 info[stripe].count++;
 733         }
 734
 735         /* alloc and initialize lov request */
 736         shift = 0;
 737         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
 738                 struct lov_oinfo *loi = NULL;
 739                 struct lov_request *req;
 740
 741                 if (info[i].count == 0)
 742                         continue;
 743
 744                 loi = oinfo->oi_md->lsm_oinfo[i];
 745                 if (!lov_check_and_wait_active(lov, loi->loi_ost_idx)) {
 746                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 747                         GOTO(out, rc = -EIO);
 748                 }
 749
 750                 OBD_ALLOC(req, sizeof(*req));
 751                 if (req == NULL)
 752                         GOTO(out, rc = -ENOMEM);
 753
 754                 OBDO_ALLOC(req->rq_oi.oi_oa);
 755                 if (req->rq_oi.oi_oa == NULL) {
 756                         OBD_FREE(req, sizeof(*req));
 757                         GOTO(out, rc = -ENOMEM);
 758                 }
 759
 760                 if (oinfo->oi_oa) {
 761                         memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
 762                                sizeof(*req->rq_oi.oi_oa));
 763                 }
 764                 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
 765                 req->rq_oi.oi_oa->o_stripe_idx = i;
 766
 767                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 768                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 769                 if (req->rq_oi.oi_md == NULL) {
 770                         OBDO_FREE(req->rq_oi.oi_oa);
 771                         OBD_FREE(req, sizeof(*req));
 772                         GOTO(out, rc = -ENOMEM);
 773                 }
 774
 775                 req->rq_idx = loi->loi_ost_idx;
 776                 req->rq_stripe = i;
 777
 778                 /* XXX LOV STACKING */
 779                 req->rq_oi.oi_md->lsm_oi = loi->loi_oi;
 780                 req->rq_oabufs = info[i].count;
 781                 req->rq_pgaidx = shift;
 782                 shift += req->rq_oabufs;
 783
 784                 /* remember the index for sort brw_page array */
 785                 info[i].index = req->rq_pgaidx;
 786
 787                 req->rq_oi.oi_capa = oinfo->oi_capa;
 788
 789                 lov_set_add_req(req, set);
 790         }
 791         if (!set->set_count)
 792                 GOTO(out, rc = -EIO);
 793
 794         /* rotate & sort the brw_page array */
 795         for (i = 0; i < oa_bufs; i++) {
 796                 int stripe = lov_stripe_number(oinfo->oi_md, pga[i].off);
 797
 798                 shift = info[stripe].index + info[stripe].off;
 799                 LASSERT(shift < oa_bufs);
 800                 set->set_pga[shift] = pga[i];
 801                 lov_stripe_offset(oinfo->oi_md, pga[i].off, stripe,
 802                                   &set->set_pga[shift].off);
 803                 info[stripe].off++;
 804         }
 805 out:
 806         if (info)
 807                 OBD_FREE_LARGE(info,
 808                                sizeof(*info) * oinfo->oi_md->lsm_stripe_count);
 809
 810         if (rc == 0)
 811                 *reqset = set;
 812         else
 813                 lov_fini_brw_set(set);
 814
 815         return rc;
 816 }
 817
 818 int lov_fini_getattr_set(struct lov_request_set *set)
 819 {
 820         int rc = 0;
 821
 822         if (set == NULL)
 823                 return 0;
 824         LASSERT(set->set_exp);
 825         if (atomic_read(&set->set_completes))
 826                 rc = common_attr_done(set);
 827
 828         lov_put_reqset(set);
 829
 830         return rc;
 831 }
 832
 833 /* The callback for osc_getattr_async that finalizes a request info when a
 834  * response is received. */
 835 static int cb_getattr_update(void *cookie, int rc)
 836 {
 837         struct obd_info *oinfo = cookie;
 838         struct lov_request *lovreq;
 839
 840         lovreq = container_of(oinfo, struct lov_request, rq_oi);
 841         return lov_update_common_set(lovreq->rq_rqset, lovreq, rc);
 842 }
 843
 844 int lov_prep_getattr_set(struct obd_export *exp, struct obd_info *oinfo,
 845                          struct lov_request_set **reqset)
 846 {
 847         struct lov_request_set *set;
 848         struct lov_obd *lov = &exp->exp_obd->u.lov;
 849         int rc = 0, i;
 850
 851         OBD_ALLOC(set, sizeof(*set));
 852         if (set == NULL)
 853                 return -ENOMEM;
 854         lov_init_set(set);
 855
 856         set->set_exp = exp;
 857         set->set_oi = oinfo;
 858
 859         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
 860                 struct lov_oinfo *loi;
 861                 struct lov_request *req;
 862
 863                 loi = oinfo->oi_md->lsm_oinfo[i];
 864                 if (!lov_check_and_wait_active(lov, loi->loi_ost_idx)) {
 865                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 866                         if (oinfo->oi_oa->o_valid & OBD_MD_FLEPOCH)
 867                                 /* SOM requires all the OSTs to be active. */
 868                                 GOTO(out_set, rc = -EIO);
 869                         continue;
 870                 }
 871
 872                 OBD_ALLOC(req, sizeof(*req));
 873                 if (req == NULL)
 874                         GOTO(out_set, rc = -ENOMEM);
 875
 876                 req->rq_stripe = i;
 877                 req->rq_idx = loi->loi_ost_idx;
 878
 879                 OBDO_ALLOC(req->rq_oi.oi_oa);
 880                 if (req->rq_oi.oi_oa == NULL) {
 881                         OBD_FREE(req, sizeof(*req));
 882                         GOTO(out_set, rc = -ENOMEM);
 883                 }
 884                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
 885                        sizeof(*req->rq_oi.oi_oa));
 886                 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
 887                 req->rq_oi.oi_cb_up = cb_getattr_update;
 888                 req->rq_oi.oi_capa = oinfo->oi_capa;
 889
 890                 lov_set_add_req(req, set);
 891         }
 892         if (!set->set_count)
 893                 GOTO(out_set, rc = -EIO);
 894         *reqset = set;
 895         return rc;
 896 out_set:
 897         lov_fini_getattr_set(set);
 898         return rc;
 899 }
 900
 901 int lov_fini_destroy_set(struct lov_request_set *set)
 902 {
 903         if (set == NULL)
 904                 return 0;
 905         LASSERT(set->set_exp);
 906         if (atomic_read(&set->set_completes)) {
 907                 /* FIXME update qos data here */
 908         }
 909
 910         lov_put_reqset(set);
 911
 912         return 0;
 913 }
 914
 915 int lov_prep_destroy_set(struct obd_export *exp, struct obd_info *oinfo,
 916                          struct obdo *src_oa, struct lov_stripe_md *lsm,
 917                          struct obd_trans_info *oti,
 918                          struct lov_request_set **reqset)
 919 {
 920         struct lov_request_set *set;
 921         struct lov_obd *lov = &exp->exp_obd->u.lov;
 922         int rc = 0, i;
 923
 924         OBD_ALLOC(set, sizeof(*set));
 925         if (set == NULL)
 926                 return -ENOMEM;
 927         lov_init_set(set);
 928
 929         set->set_exp = exp;
 930         set->set_oi = oinfo;
 931         set->set_oi->oi_md = lsm;
 932         set->set_oi->oi_oa = src_oa;
 933         set->set_oti = oti;
 934         if (oti != NULL && src_oa->o_valid & OBD_MD_FLCOOKIE)
 935                 set->set_cookies = oti->oti_logcookies;
 936
 937         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 938                 struct lov_oinfo *loi;
 939                 struct lov_request *req;
 940
 941                 loi = lsm->lsm_oinfo[i];
 942                 if (!lov_check_and_wait_active(lov, loi->loi_ost_idx)) {
 943                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 944                         continue;
 945                 }
 946
 947                 OBD_ALLOC(req, sizeof(*req));
 948                 if (req == NULL)
 949                         GOTO(out_set, rc = -ENOMEM);
 950
 951                 req->rq_stripe = i;
 952                 req->rq_idx = loi->loi_ost_idx;
 953
 954                 OBDO_ALLOC(req->rq_oi.oi_oa);
 955                 if (req->rq_oi.oi_oa == NULL) {
 956                         OBD_FREE(req, sizeof(*req));
 957                         GOTO(out_set, rc = -ENOMEM);
 958                 }
 959                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
 960                 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
 961                 lov_set_add_req(req, set);
 962         }
 963         if (!set->set_count)
 964                 GOTO(out_set, rc = -EIO);
 965         *reqset = set;
 966         return rc;
 967 out_set:
 968         lov_fini_destroy_set(set);
 969         return rc;
 970 }
 971
 972 int lov_fini_setattr_set(struct lov_request_set *set)
 973 {
 974         int rc = 0;
 975
 976         if (set == NULL)
 977                 return 0;
 978         LASSERT(set->set_exp);
 979         if (atomic_read(&set->set_completes)) {
 980                 rc = common_attr_done(set);
 981                 /* FIXME update qos data here */
 982         }
 983
 984         lov_put_reqset(set);
 985         return rc;
 986 }
 987
 988 int lov_update_setattr_set(struct lov_request_set *set,
 989                            struct lov_request *req, int rc)
 990 {
 991         struct lov_obd *lov = &req->rq_rqset->set_exp->exp_obd->u.lov;
 992         struct lov_stripe_md *lsm = req->rq_rqset->set_oi->oi_md;
 993
 994         lov_update_set(set, req, rc);
 995
 996         /* grace error on inactive ost */
 997         if (rc && !(lov->lov_tgts[req->rq_idx] &&
 998                     lov->lov_tgts[req->rq_idx]->ltd_active))
 999                 rc = 0;
1000
1001         if (rc == 0) {
1002                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLCTIME)
1003                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_ctime =
1004                                 req->rq_oi.oi_oa->o_ctime;
1005                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLMTIME)
1006                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_mtime =
1007                                 req->rq_oi.oi_oa->o_mtime;
1008                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLATIME)
1009                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_atime =
1010                                 req->rq_oi.oi_oa->o_atime;
1011         }
1012
1013         return rc;
1014 }
1015
1016 /* The callback for osc_setattr_async that finalizes a request info when a
1017  * response is received. */
1018 static int cb_setattr_update(void *cookie, int rc)
1019 {
1020         struct obd_info *oinfo = cookie;
1021         struct lov_request *lovreq;
1022
1023         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1024         return lov_update_setattr_set(lovreq->rq_rqset, lovreq, rc);
1025 }
1026
1027 int lov_prep_setattr_set(struct obd_export *exp, struct obd_info *oinfo,
1028                          struct obd_trans_info *oti,
1029                          struct lov_request_set **reqset)
1030 {
1031         struct lov_request_set *set;
1032         struct lov_obd *lov = &exp->exp_obd->u.lov;
1033         int rc = 0, i;
1034
1035         OBD_ALLOC(set, sizeof(*set));
1036         if (set == NULL)
1037                 return -ENOMEM;
1038         lov_init_set(set);
1039
1040         set->set_exp = exp;
1041         set->set_oti = oti;
1042         set->set_oi = oinfo;
1043         if (oti != NULL && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
1044                 set->set_cookies = oti->oti_logcookies;
1045
1046         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1047                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1048                 struct lov_request *req;
1049
1050                 if (!lov_check_and_wait_active(lov, loi->loi_ost_idx)) {
1051                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1052                         continue;
1053                 }
1054
1055                 OBD_ALLOC(req, sizeof(*req));
1056                 if (req == NULL)
1057                         GOTO(out_set, rc = -ENOMEM);
1058                 req->rq_stripe = i;
1059                 req->rq_idx = loi->loi_ost_idx;
1060
1061                 OBDO_ALLOC(req->rq_oi.oi_oa);
1062                 if (req->rq_oi.oi_oa == NULL) {
1063                         OBD_FREE(req, sizeof(*req));
1064                         GOTO(out_set, rc = -ENOMEM);
1065                 }
1066                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1067                        sizeof(*req->rq_oi.oi_oa));
1068                 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
1069                 req->rq_oi.oi_oa->o_stripe_idx = i;
1070                 req->rq_oi.oi_cb_up = cb_setattr_update;
1071                 req->rq_oi.oi_capa = oinfo->oi_capa;
1072
1073                 if (oinfo->oi_oa->o_valid & OBD_MD_FLSIZE) {
1074                         int off = lov_stripe_offset(oinfo->oi_md,
1075                                                     oinfo->oi_oa->o_size, i,
1076                                                     &req->rq_oi.oi_oa->o_size);
1077
1078                         if (off < 0 && req->rq_oi.oi_oa->o_size)
1079                                 req->rq_oi.oi_oa->o_size--;
1080
1081                         CDEBUG(D_INODE, "stripe %d has size %llu/%llu\n",
1082                                i, req->rq_oi.oi_oa->o_size,
1083                                oinfo->oi_oa->o_size);
1084                 }
1085                 lov_set_add_req(req, set);
1086         }
1087         if (!set->set_count)
1088                 GOTO(out_set, rc = -EIO);
1089         *reqset = set;
1090         return rc;
1091 out_set:
1092         lov_fini_setattr_set(set);
1093         return rc;
1094 }
1095
1096 int lov_fini_punch_set(struct lov_request_set *set)
1097 {
1098         int rc = 0;
1099
1100         if (set == NULL)
1101                 return 0;
1102         LASSERT(set->set_exp);
1103         if (atomic_read(&set->set_completes)) {
1104                 rc = -EIO;
1105                 /* FIXME update qos data here */
1106                 if (atomic_read(&set->set_success))
1107                         rc = common_attr_done(set);
1108         }
1109
1110         lov_put_reqset(set);
1111
1112         return rc;
1113 }
1114
1115 int lov_update_punch_set(struct lov_request_set *set,
1116                          struct lov_request *req, int rc)
1117 {
1118         struct lov_obd *lov = &req->rq_rqset->set_exp->exp_obd->u.lov;
1119         struct lov_stripe_md *lsm = req->rq_rqset->set_oi->oi_md;
1120
1121         lov_update_set(set, req, rc);
1122
1123         /* grace error on inactive ost */
1124         if (rc && !lov->lov_tgts[req->rq_idx]->ltd_active)
1125                 rc = 0;
1126
1127         if (rc == 0) {
1128                 lov_stripe_lock(lsm);
1129                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLBLOCKS) {
1130                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_blocks =
1131                                 req->rq_oi.oi_oa->o_blocks;
1132                 }
1133
1134                 lov_stripe_unlock(lsm);
1135         }
1136
1137         return rc;
1138 }
1139
1140 /* The callback for osc_punch that finalizes a request info when a response
1141  * is received. */
1142 static int cb_update_punch(void *cookie, int rc)
1143 {
1144         struct obd_info *oinfo = cookie;
1145         struct lov_request *lovreq;
1146
1147         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1148         return lov_update_punch_set(lovreq->rq_rqset, lovreq, rc);
1149 }
1150
1151 int lov_prep_punch_set(struct obd_export *exp, struct obd_info *oinfo,
1152                        struct obd_trans_info *oti,
1153                        struct lov_request_set **reqset)
1154 {
1155         struct lov_request_set *set;
1156         struct lov_obd *lov = &exp->exp_obd->u.lov;
1157         int rc = 0, i;
1158
1159         OBD_ALLOC(set, sizeof(*set));
1160         if (set == NULL)
1161                 return -ENOMEM;
1162         lov_init_set(set);
1163
1164         set->set_oi = oinfo;
1165         set->set_exp = exp;
1166
1167         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1168                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1169                 struct lov_request *req;
1170                 obd_off rs, re;
1171
1172                 if (!lov_stripe_intersects(oinfo->oi_md, i,
1173                                            oinfo->oi_policy.l_extent.start,
1174                                            oinfo->oi_policy.l_extent.end,
1175                                            &rs, &re))
1176                         continue;
1177
1178                 if (!lov_check_and_wait_active(lov, loi->loi_ost_idx)) {
1179                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1180                         GOTO(out_set, rc = -EIO);
1181                 }
1182
1183                 OBD_ALLOC(req, sizeof(*req));
1184                 if (req == NULL)
1185                         GOTO(out_set, rc = -ENOMEM);
1186                 req->rq_stripe = i;
1187                 req->rq_idx = loi->loi_ost_idx;
1188
1189                 OBDO_ALLOC(req->rq_oi.oi_oa);
1190                 if (req->rq_oi.oi_oa == NULL) {
1191                         OBD_FREE(req, sizeof(*req));
1192                         GOTO(out_set, rc = -ENOMEM);
1193                 }
1194                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1195                        sizeof(*req->rq_oi.oi_oa));
1196                 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
1197                 req->rq_oi.oi_oa->o_valid |= OBD_MD_FLGROUP;
1198
1199                 req->rq_oi.oi_oa->o_stripe_idx = i;
1200                 req->rq_oi.oi_cb_up = cb_update_punch;
1201
1202                 req->rq_oi.oi_policy.l_extent.start = rs;
1203                 req->rq_oi.oi_policy.l_extent.end = re;
1204                 req->rq_oi.oi_policy.l_extent.gid = -1;
1205
1206                 req->rq_oi.oi_capa = oinfo->oi_capa;
1207
1208                 lov_set_add_req(req, set);
1209         }
1210         if (!set->set_count)
1211                 GOTO(out_set, rc = -EIO);
1212         *reqset = set;
1213         return rc;
1214 out_set:
1215         lov_fini_punch_set(set);
1216         return rc;
1217 }
1218
1219 int lov_fini_sync_set(struct lov_request_set *set)
1220 {
1221         int rc = 0;
1222
1223         if (set == NULL)
1224                 return 0;
1225         LASSERT(set->set_exp);
1226         if (atomic_read(&set->set_completes)) {
1227                 if (!atomic_read(&set->set_success))
1228                         rc = -EIO;
1229                 /* FIXME update qos data here */
1230         }
1231
1232         lov_put_reqset(set);
1233
1234         return rc;
1235 }
1236
1237 /* The callback for osc_sync that finalizes a request info when a
1238  * response is received. */
1239 static int cb_sync_update(void *cookie, int rc)
1240 {
1241         struct obd_info *oinfo = cookie;
1242         struct lov_request *lovreq;
1243
1244         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1245         return lov_update_common_set(lovreq->rq_rqset, lovreq, rc);
1246 }
1247
1248 int lov_prep_sync_set(struct obd_export *exp, struct obd_info *oinfo,
1249                       obd_off start, obd_off end,
1250                       struct lov_request_set **reqset)
1251 {
1252         struct lov_request_set *set;
1253         struct lov_obd *lov = &exp->exp_obd->u.lov;
1254         int rc = 0, i;
1255
1256         OBD_ALLOC_PTR(set);
1257         if (set == NULL)
1258                 return -ENOMEM;
1259         lov_init_set(set);
1260
1261         set->set_exp = exp;
1262         set->set_oi = oinfo;
1263
1264         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1265                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1266                 struct lov_request *req;
1267                 obd_off rs, re;
1268
1269                 if (!lov_check_and_wait_active(lov, loi->loi_ost_idx)) {
1270                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1271                         continue;
1272                 }
1273
1274                 if (!lov_stripe_intersects(oinfo->oi_md, i, start, end, &rs,
1275                                            &re))
1276                         continue;
1277
1278                 OBD_ALLOC_PTR(req);
1279                 if (req == NULL)
1280                         GOTO(out_set, rc = -ENOMEM);
1281                 req->rq_stripe = i;
1282                 req->rq_idx = loi->loi_ost_idx;
1283
1284                 OBDO_ALLOC(req->rq_oi.oi_oa);
1285                 if (req->rq_oi.oi_oa == NULL) {
1286                         OBD_FREE(req, sizeof(*req));
1287                         GOTO(out_set, rc = -ENOMEM);
1288                 }
1289                 *req->rq_oi.oi_oa = *oinfo->oi_oa;
1290                 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
1291                 req->rq_oi.oi_oa->o_stripe_idx = i;
1292
1293                 req->rq_oi.oi_policy.l_extent.start = rs;
1294                 req->rq_oi.oi_policy.l_extent.end = re;
1295                 req->rq_oi.oi_policy.l_extent.gid = -1;
1296                 req->rq_oi.oi_cb_up = cb_sync_update;
1297
1298                 lov_set_add_req(req, set);
1299         }
1300         if (!set->set_count)
1301                 GOTO(out_set, rc = -EIO);
1302         *reqset = set;
1303         return rc;
1304 out_set:
1305         lov_fini_sync_set(set);
1306         return rc;
1307 }
1308
1309 #define LOV_U64_MAX ((__u64)~0ULL)
1310 #define LOV_SUM_MAX(tot, add)                                      \
1311         do {                                                        \
1312                 if ((tot) + (add) < (tot))                            \
1313                         (tot) = LOV_U64_MAX;                        \
1314                 else                                                \
1315                         (tot) += (add);                          \
1316         } while (0)
1317
1318 int lov_fini_statfs(struct obd_device *obd, struct obd_statfs *osfs,int success)
1319 {
1320         if (success) {
1321                 __u32 expected_stripes = lov_get_stripecnt(&obd->u.lov,
1322                                                            LOV_MAGIC, 0);
1323                 if (osfs->os_files != LOV_U64_MAX)
1324                         lov_do_div64(osfs->os_files, expected_stripes);
1325                 if (osfs->os_ffree != LOV_U64_MAX)
1326                         lov_do_div64(osfs->os_ffree, expected_stripes);
1327
1328                 spin_lock(&obd->obd_osfs_lock);
1329                 memcpy(&obd->obd_osfs, osfs, sizeof(*osfs));
1330                 obd->obd_osfs_age = cfs_time_current_64();
1331                 spin_unlock(&obd->obd_osfs_lock);
1332                 return 0;
1333         }
1334
1335         return -EIO;
1336 }
1337
1338 int lov_fini_statfs_set(struct lov_request_set *set)
1339 {
1340         int rc = 0;
1341
1342         if (set == NULL)
1343                 return 0;
1344
1345         if (atomic_read(&set->set_completes)) {
1346                 rc = lov_fini_statfs(set->set_obd, set->set_oi->oi_osfs,
1347                                      atomic_read(&set->set_success));
1348         }
1349         lov_put_reqset(set);
1350         return rc;
1351 }
1352
1353 void lov_update_statfs(struct obd_statfs *osfs, struct obd_statfs *lov_sfs,
1354                        int success)
1355 {
1356         int shift = 0, quit = 0;
1357         __u64 tmp;
1358
1359         if (success == 0) {
1360                 memcpy(osfs, lov_sfs, sizeof(*lov_sfs));
1361         } else {
1362                 if (osfs->os_bsize != lov_sfs->os_bsize) {
1363                         /* assume all block sizes are always powers of 2 */
1364                         /* get the bits difference */
1365                         tmp = osfs->os_bsize | lov_sfs->os_bsize;
1366                         for (shift = 0; shift <= 64; ++shift) {
1367                                 if (tmp & 1) {
1368                                         if (quit)
1369                                                 break;
1370                                         else
1371                                                 quit = 1;
1372                                         shift = 0;
1373                                 }
1374                                 tmp >>= 1;
1375                         }
1376                 }
1377
1378                 if (osfs->os_bsize < lov_sfs->os_bsize) {
1379                         osfs->os_bsize = lov_sfs->os_bsize;
1380
1381                         osfs->os_bfree  >>= shift;
1382                         osfs->os_bavail >>= shift;
1383                         osfs->os_blocks >>= shift;
1384                 } else if (shift != 0) {
1385                         lov_sfs->os_bfree  >>= shift;
1386                         lov_sfs->os_bavail >>= shift;
1387                         lov_sfs->os_blocks >>= shift;
1388                 }
1389                 osfs->os_bfree += lov_sfs->os_bfree;
1390                 osfs->os_bavail += lov_sfs->os_bavail;
1391                 osfs->os_blocks += lov_sfs->os_blocks;
1392                 /* XXX not sure about this one - depends on policy.
1393                  *   - could be minimum if we always stripe on all OBDs
1394                  *     (but that would be wrong for any other policy,
1395                  *     if one of the OBDs has no more objects left)
1396                  *   - could be sum if we stripe whole objects
1397                  *   - could be average, just to give a nice number
1398                  *
1399                  * To give a "reasonable" (if not wholly accurate)
1400                  * number, we divide the total number of free objects
1401                  * by expected stripe count (watch out for overflow).
1402                  */
1403                 LOV_SUM_MAX(osfs->os_files, lov_sfs->os_files);
1404                 LOV_SUM_MAX(osfs->os_ffree, lov_sfs->os_ffree);
1405         }
1406 }
1407
1408 /* The callback for osc_statfs_async that finalizes a request info when a
1409  * response is received. */
1410 static int cb_statfs_update(void *cookie, int rc)
1411 {
1412         struct obd_info *oinfo = cookie;
1413         struct lov_request *lovreq;
1414         struct lov_request_set *set;
1415         struct obd_statfs *osfs, *lov_sfs;
1416         struct lov_obd *lov;
1417         struct lov_tgt_desc *tgt;
1418         struct obd_device *lovobd, *tgtobd;
1419         int success;
1420
1421         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1422         set = lovreq->rq_rqset;
1423         lovobd = set->set_obd;
1424         lov = &lovobd->u.lov;
1425         osfs = set->set_oi->oi_osfs;
1426         lov_sfs = oinfo->oi_osfs;
1427         success = atomic_read(&set->set_success);
1428         /* XXX: the same is done in lov_update_common_set, however
1429            lovset->set_exp is not initialized. */
1430         lov_update_set(set, lovreq, rc);
1431         if (rc)
1432                 GOTO(out, rc);
1433
1434         obd_getref(lovobd);
1435         tgt = lov->lov_tgts[lovreq->rq_idx];
1436         if (!tgt || !tgt->ltd_active)
1437                 GOTO(out_update, rc);
1438
1439         tgtobd = class_exp2obd(tgt->ltd_exp);
1440         spin_lock(&tgtobd->obd_osfs_lock);
1441         memcpy(&tgtobd->obd_osfs, lov_sfs, sizeof(*lov_sfs));
1442         if ((oinfo->oi_flags & OBD_STATFS_FROM_CACHE) == 0)
1443                 tgtobd->obd_osfs_age = cfs_time_current_64();
1444         spin_unlock(&tgtobd->obd_osfs_lock);
1445
1446 out_update:
1447         lov_update_statfs(osfs, lov_sfs, success);
1448         obd_putref(lovobd);
1449
1450 out:
1451         if (set->set_oi->oi_flags & OBD_STATFS_PTLRPCD &&
1452             lov_set_finished(set, 0)) {
1453                 lov_statfs_interpret(NULL, set, set->set_count !=
1454                                      atomic_read(&set->set_success));
1455         }
1456
1457         return 0;
1458 }
1459
1460 int lov_prep_statfs_set(struct obd_device *obd, struct obd_info *oinfo,
1461                         struct lov_request_set **reqset)
1462 {
1463         struct lov_request_set *set;
1464         struct lov_obd *lov = &obd->u.lov;
1465         int rc = 0, i;
1466
1467         OBD_ALLOC(set, sizeof(*set));
1468         if (set == NULL)
1469                 return -ENOMEM;
1470         lov_init_set(set);
1471
1472         set->set_obd = obd;
1473         set->set_oi = oinfo;
1474
1475         /* We only get block data from the OBD */
1476         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1477                 struct lov_request *req;
1478
1479                 if (lov->lov_tgts[i] == NULL ||
1480                     (!lov_check_and_wait_active(lov, i) &&
1481                      (oinfo->oi_flags & OBD_STATFS_NODELAY))) {
1482                         CDEBUG(D_HA, "lov idx %d inactive\n", i);
1483                         continue;
1484                 }
1485
1486                 /* skip targets that have been explicitly disabled by the
1487                  * administrator */
1488                 if (!lov->lov_tgts[i]->ltd_exp) {
1489                         CDEBUG(D_HA, "lov idx %d administratively disabled\n", i);
1490                         continue;
1491                 }
1492
1493                 OBD_ALLOC(req, sizeof(*req));
1494                 if (req == NULL)
1495                         GOTO(out_set, rc = -ENOMEM);
1496
1497                 OBD_ALLOC(req->rq_oi.oi_osfs, sizeof(*req->rq_oi.oi_osfs));
1498                 if (req->rq_oi.oi_osfs == NULL) {
1499                         OBD_FREE(req, sizeof(*req));
1500                         GOTO(out_set, rc = -ENOMEM);
1501                 }
1502
1503                 req->rq_idx = i;
1504                 req->rq_oi.oi_cb_up = cb_statfs_update;
1505                 req->rq_oi.oi_flags = oinfo->oi_flags;
1506
1507                 lov_set_add_req(req, set);
1508         }
1509         if (!set->set_count)
1510                 GOTO(out_set, rc = -EIO);
1511         *reqset = set;
1512         return rc;
1513 out_set:
1514         lov_fini_statfs_set(set);
1515         return rc;
1516 }