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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  */
  38
  39 #ifndef EXPORT_SYMTAB
  40 # define EXPORT_SYMTAB
  41 #endif
  42 #define DEBUG_SUBSYSTEM S_LOV
  43
  44 #ifdef __KERNEL__
  45 #include <libcfs/libcfs.h>
  46 #else
  47 #include <liblustre.h>
  48 #endif
  49
  50 #include <obd_class.h>
  51 #include <obd_lov.h>
  52 #include <lustre/lustre_idl.h>
  53
  54 #include "lov_internal.h"
  55
  56 static void lov_init_set(struct lov_request_set *set)
  57 {
  58         set->set_count = 0;
  59         cfs_atomic_set(&set->set_completes, 0);
  60         cfs_atomic_set(&set->set_success, 0);
  61         set->set_cookies = 0;
  62         CFS_INIT_LIST_HEAD(&set->set_list);
  63         cfs_atomic_set(&set->set_refcount, 1);
  64         cfs_waitq_init(&set->set_waitq);
  65         cfs_spin_lock_init(&set->set_lock);
  66 }
  67
  68 void lov_finish_set(struct lov_request_set *set)
  69 {
  70         cfs_list_t *pos, *n;
  71         ENTRY;
  72
  73         LASSERT(set);
  74         cfs_list_for_each_safe(pos, n, &set->set_list) {
  75                 struct lov_request *req = cfs_list_entry(pos,
  76                                                          struct lov_request,
  77                                                          rq_link);
  78                 cfs_list_del_init(&req->rq_link);
  79
  80                 if (req->rq_oi.oi_oa)
  81                         OBDO_FREE(req->rq_oi.oi_oa);
  82                 if (req->rq_oi.oi_md)
  83                         OBD_FREE_LARGE(req->rq_oi.oi_md, req->rq_buflen);
  84                 if (req->rq_oi.oi_osfs)
  85                         OBD_FREE(req->rq_oi.oi_osfs,
  86                                  sizeof(*req->rq_oi.oi_osfs));
  87                 OBD_FREE(req, sizeof(*req));
  88         }
  89
  90         if (set->set_pga) {
  91                 int len = set->set_oabufs * sizeof(*set->set_pga);
  92                 OBD_FREE_LARGE(set->set_pga, len);
  93         }
  94         if (set->set_lockh)
  95                 lov_llh_put(set->set_lockh);
  96
  97         OBD_FREE(set, sizeof(*set));
  98         EXIT;
  99 }
 100
 101 int lov_finished_set(struct lov_request_set *set)
 102 {
 103         int completes = cfs_atomic_read(&set->set_completes);
 104
 105         CDEBUG(D_INFO, "check set %d/%d\n", completes,
 106                set->set_count);
 107         return completes == set->set_count;
 108 }
 109
 110 void lov_update_set(struct lov_request_set *set,
 111                     struct lov_request *req, int rc)
 112 {
 113         req->rq_complete = 1;
 114         req->rq_rc = rc;
 115
 116         cfs_atomic_inc(&set->set_completes);
 117         if (rc == 0)
 118                 cfs_atomic_inc(&set->set_success);
 119
 120         cfs_waitq_signal(&set->set_waitq);
 121 }
 122
 123 int lov_update_common_set(struct lov_request_set *set,
 124                           struct lov_request *req, int rc)
 125 {
 126         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 127         ENTRY;
 128
 129         lov_update_set(set, req, rc);
 130
 131         /* grace error on inactive ost */
 132         if (rc && !(lov->lov_tgts[req->rq_idx] &&
 133                     lov->lov_tgts[req->rq_idx]->ltd_active))
 134                 rc = 0;
 135
 136         /* FIXME in raid1 regime, should return 0 */
 137         RETURN(rc);
 138 }
 139
 140 void lov_set_add_req(struct lov_request *req, struct lov_request_set *set)
 141 {
 142         cfs_list_add_tail(&req->rq_link, &set->set_list);
 143         set->set_count++;
 144         req->rq_rqset = set;
 145 }
 146
 147 extern void osc_update_enqueue(struct lustre_handle *lov_lockhp,
 148                                struct lov_oinfo *loi, int flags,
 149                                struct ost_lvb *lvb, __u32 mode, int rc);
 150
 151 static int lov_update_enqueue_lov(struct obd_export *exp,
 152                                   struct lustre_handle *lov_lockhp,
 153                                   struct lov_oinfo *loi, int flags, int idx,
 154                                   __u64 oid, int rc)
 155 {
 156         struct lov_obd *lov = &exp->exp_obd->u.lov;
 157
 158         if (rc != ELDLM_OK &&
 159             !(rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT))) {
 160                 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
 161                 if (lov->lov_tgts[idx] && lov->lov_tgts[idx]->ltd_active) {
 162                         /* -EUSERS used by OST to report file contention */
 163                         if (rc != -EINTR && rc != -EUSERS)
 164                                 CERROR("enqueue objid "LPX64" subobj "
 165                                        LPX64" on OST idx %d: rc %d\n",
 166                                        oid, loi->loi_id, loi->loi_ost_idx, rc);
 167                 } else
 168                         rc = ELDLM_OK;
 169         }
 170         return rc;
 171 }
 172
 173 int lov_update_enqueue_set(struct lov_request *req, __u32 mode, int rc)
 174 {
 175         struct lov_request_set *set = req->rq_rqset;
 176         struct lustre_handle *lov_lockhp;
 177         struct obd_info *oi = set->set_oi;
 178         struct lov_oinfo *loi;
 179         ENTRY;
 180
 181         LASSERT(oi != NULL);
 182
 183         lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
 184         loi = oi->oi_md->lsm_oinfo[req->rq_stripe];
 185
 186         /* XXX LOV STACKING: OSC gets a copy, created in lov_prep_enqueue_set
 187          * and that copy can be arbitrarily out of date.
 188          *
 189          * The LOV API is due for a serious rewriting anyways, and this
 190          * can be addressed then. */
 191
 192         lov_stripe_lock(oi->oi_md);
 193         osc_update_enqueue(lov_lockhp, loi, oi->oi_flags,
 194                            &req->rq_oi.oi_md->lsm_oinfo[0]->loi_lvb, mode, rc);
 195         if (rc == ELDLM_LOCK_ABORTED && (oi->oi_flags & LDLM_FL_HAS_INTENT))
 196                 memset(lov_lockhp, 0, sizeof *lov_lockhp);
 197         rc = lov_update_enqueue_lov(set->set_exp, lov_lockhp, loi, oi->oi_flags,
 198                                     req->rq_idx, oi->oi_md->lsm_object_id, rc);
 199         lov_stripe_unlock(oi->oi_md);
 200         lov_update_set(set, req, rc);
 201         RETURN(rc);
 202 }
 203
 204 /* The callback for osc_enqueue that updates lov info for every OSC request. */
 205 static int cb_update_enqueue(void *cookie, int rc)
 206 {
 207         struct obd_info *oinfo = cookie;
 208         struct ldlm_enqueue_info *einfo;
 209         struct lov_request *lovreq;
 210
 211         lovreq = container_of(oinfo, struct lov_request, rq_oi);
 212         einfo = lovreq->rq_rqset->set_ei;
 213         return lov_update_enqueue_set(lovreq, einfo->ei_mode, rc);
 214 }
 215
 216 static int enqueue_done(struct lov_request_set *set, __u32 mode)
 217 {
 218         struct lov_request *req;
 219         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 220         int completes = cfs_atomic_read(&set->set_completes);
 221         int rc = 0;
 222         ENTRY;
 223
 224         /* enqueue/match success, just return */
 225         if (completes && completes == cfs_atomic_read(&set->set_success))
 226                 RETURN(0);
 227
 228         /* cancel enqueued/matched locks */
 229         cfs_list_for_each_entry(req, &set->set_list, rq_link) {
 230                 struct lustre_handle *lov_lockhp;
 231
 232                 if (!req->rq_complete || req->rq_rc)
 233                         continue;
 234
 235                 lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
 236                 LASSERT(lov_lockhp);
 237                 if (!lustre_handle_is_used(lov_lockhp))
 238                         continue;
 239
 240                 rc = obd_cancel(lov->lov_tgts[req->rq_idx]->ltd_exp,
 241                                 req->rq_oi.oi_md, mode, lov_lockhp);
 242                 if (rc && lov->lov_tgts[req->rq_idx] &&
 243                     lov->lov_tgts[req->rq_idx]->ltd_active)
 244                         CERROR("cancelling obdjid "LPX64" on OST "
 245                                "idx %d error: rc = %d\n",
 246                                req->rq_oi.oi_md->lsm_object_id,
 247                                req->rq_idx, rc);
 248         }
 249         if (set->set_lockh)
 250                 lov_llh_put(set->set_lockh);
 251         RETURN(rc);
 252 }
 253
 254 int lov_fini_enqueue_set(struct lov_request_set *set, __u32 mode, int rc,
 255                          struct ptlrpc_request_set *rqset)
 256 {
 257         int ret = 0;
 258         ENTRY;
 259
 260         if (set == NULL)
 261                 RETURN(0);
 262         LASSERT(set->set_exp);
 263         /* Do enqueue_done only for sync requests and if any request
 264          * succeeded. */
 265         if (!rqset) {
 266                 if (rc)
 267                         cfs_atomic_set(&set->set_completes, 0);
 268                 ret = enqueue_done(set, mode);
 269         } else if (set->set_lockh)
 270                 lov_llh_put(set->set_lockh);
 271
 272         lov_put_reqset(set);
 273
 274         RETURN(rc ? rc : ret);
 275 }
 276
 277 int lov_prep_enqueue_set(struct obd_export *exp, struct obd_info *oinfo,
 278                          struct ldlm_enqueue_info *einfo,
 279                          struct lov_request_set **reqset)
 280 {
 281         struct lov_obd *lov = &exp->exp_obd->u.lov;
 282         struct lov_request_set *set;
 283         int i, rc = 0;
 284         ENTRY;
 285
 286         OBD_ALLOC(set, sizeof(*set));
 287         if (set == NULL)
 288                 RETURN(-ENOMEM);
 289         lov_init_set(set);
 290
 291         set->set_exp = exp;
 292         set->set_oi = oinfo;
 293         set->set_ei = einfo;
 294         set->set_lockh = lov_llh_new(oinfo->oi_md);
 295         if (set->set_lockh == NULL)
 296                 GOTO(out_set, rc = -ENOMEM);
 297         oinfo->oi_lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 298
 299         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
 300                 struct lov_oinfo *loi;
 301                 struct lov_request *req;
 302                 obd_off start, end;
 303
 304                 loi = oinfo->oi_md->lsm_oinfo[i];
 305                 if (!lov_stripe_intersects(oinfo->oi_md, i,
 306                                            oinfo->oi_policy.l_extent.start,
 307                                            oinfo->oi_policy.l_extent.end,
 308                                            &start, &end))
 309                         continue;
 310
 311                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 312                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 313                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 314                         continue;
 315                 }
 316
 317                 OBD_ALLOC(req, sizeof(*req));
 318                 if (req == NULL)
 319                         GOTO(out_set, rc = -ENOMEM);
 320
 321                 req->rq_buflen = sizeof(*req->rq_oi.oi_md) +
 322                         sizeof(struct lov_oinfo *) +
 323                         sizeof(struct lov_oinfo);
 324                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 325                 if (req->rq_oi.oi_md == NULL) {
 326                         OBD_FREE(req, sizeof(*req));
 327                         GOTO(out_set, rc = -ENOMEM);
 328                 }
 329                 req->rq_oi.oi_md->lsm_oinfo[0] =
 330                         ((void *)req->rq_oi.oi_md) + sizeof(*req->rq_oi.oi_md) +
 331                         sizeof(struct lov_oinfo *);
 332
 333                 /* Set lov request specific parameters. */
 334                 req->rq_oi.oi_lockh = set->set_lockh->llh_handles + i;
 335                 req->rq_oi.oi_cb_up = cb_update_enqueue;
 336                 req->rq_oi.oi_flags = oinfo->oi_flags;
 337
 338                 LASSERT(req->rq_oi.oi_lockh);
 339
 340                 req->rq_oi.oi_policy.l_extent.gid =
 341                         oinfo->oi_policy.l_extent.gid;
 342                 req->rq_oi.oi_policy.l_extent.start = start;
 343                 req->rq_oi.oi_policy.l_extent.end = end;
 344
 345                 req->rq_idx = loi->loi_ost_idx;
 346                 req->rq_stripe = i;
 347
 348                 /* XXX LOV STACKING: submd should be from the subobj */
 349                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 350                 req->rq_oi.oi_md->lsm_object_seq = loi->loi_seq;
 351                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 352                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_kms_valid =
 353                         loi->loi_kms_valid;
 354                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_kms = loi->loi_kms;
 355                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_lvb = loi->loi_lvb;
 356
 357                 lov_set_add_req(req, set);
 358         }
 359         if (!set->set_count)
 360                 GOTO(out_set, rc = -EIO);
 361         *reqset = set;
 362         RETURN(0);
 363 out_set:
 364         lov_fini_enqueue_set(set, einfo->ei_mode, rc, NULL);
 365         RETURN(rc);
 366 }
 367
 368 int lov_fini_match_set(struct lov_request_set *set, __u32 mode, int flags)
 369 {
 370         int rc = 0;
 371         ENTRY;
 372
 373         if (set == NULL)
 374                 RETURN(0);
 375         LASSERT(set->set_exp);
 376         rc = enqueue_done(set, mode);
 377         if ((set->set_count == cfs_atomic_read(&set->set_success)) &&
 378             (flags & LDLM_FL_TEST_LOCK))
 379                 lov_llh_put(set->set_lockh);
 380
 381         lov_put_reqset(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_LARGE(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_seq = loi->loi_seq;
 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         lov_put_reqset(set);
 473
 474         RETURN(rc);
 475 }
 476
 477 int lov_prep_cancel_set(struct obd_export *exp, struct obd_info *oinfo,
 478                         struct lov_stripe_md *lsm, __u32 mode,
 479                         struct lustre_handle *lockh,
 480                         struct lov_request_set **reqset)
 481 {
 482         struct lov_request_set *set;
 483         int i, rc = 0;
 484         ENTRY;
 485
 486         OBD_ALLOC(set, sizeof(*set));
 487         if (set == NULL)
 488                 RETURN(-ENOMEM);
 489         lov_init_set(set);
 490
 491         set->set_exp = exp;
 492         set->set_oi = oinfo;
 493         set->set_oi->oi_md = lsm;
 494         set->set_lockh = lov_handle2llh(lockh);
 495         if (set->set_lockh == NULL) {
 496                 CERROR("LOV: invalid lov lock handle %p\n", lockh);
 497                 GOTO(out_set, rc = -EINVAL);
 498         }
 499         lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 500
 501         for (i = 0; i < lsm->lsm_stripe_count; i++){
 502                 struct lov_request *req;
 503                 struct lustre_handle *lov_lockhp;
 504                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 505
 506                 lov_lockhp = set->set_lockh->llh_handles + i;
 507                 if (!lustre_handle_is_used(lov_lockhp)) {
 508                         CDEBUG(D_INFO, "lov idx %d subobj "LPX64" no lock\n",
 509                                loi->loi_ost_idx, loi->loi_id);
 510                         continue;
 511                 }
 512
 513                 OBD_ALLOC(req, sizeof(*req));
 514                 if (req == NULL)
 515                         GOTO(out_set, rc = -ENOMEM);
 516
 517                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 518                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 519                 if (req->rq_oi.oi_md == NULL) {
 520                         OBD_FREE(req, sizeof(*req));
 521                         GOTO(out_set, rc = -ENOMEM);
 522                 }
 523
 524                 req->rq_idx = loi->loi_ost_idx;
 525                 req->rq_stripe = i;
 526
 527                 /* XXX LOV STACKING: submd should be from the subobj */
 528                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 529                 req->rq_oi.oi_md->lsm_object_seq = loi->loi_seq;
 530                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 531
 532                 lov_set_add_req(req, set);
 533         }
 534         if (!set->set_count)
 535                 GOTO(out_set, rc = -EIO);
 536         *reqset = set;
 537         RETURN(rc);
 538 out_set:
 539         lov_fini_cancel_set(set);
 540         RETURN(rc);
 541 }
 542
 543 static int lov_update_create_set(struct lov_request_set *set,
 544                                  struct lov_request *req, int rc)
 545 {
 546         struct obd_trans_info *oti = set->set_oti;
 547         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 548         struct lov_oinfo *loi;
 549         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 550         ENTRY;
 551
 552         if (rc && lov->lov_tgts[req->rq_idx] &&
 553             lov->lov_tgts[req->rq_idx]->ltd_active) {
 554                 CERROR("error creating fid "LPX64" sub-object"
 555                        " on OST idx %d/%d: rc = %d\n",
 556                        set->set_oi->oi_oa->o_id, req->rq_idx,
 557                        lsm->lsm_stripe_count, rc);
 558                 if (rc > 0) {
 559                         CERROR("obd_create returned invalid err %d\n", rc);
 560                         rc = -EIO;
 561                 }
 562         }
 563
 564         cfs_spin_lock(&set->set_lock);
 565         req->rq_stripe = cfs_atomic_read(&set->set_success);
 566         loi = lsm->lsm_oinfo[req->rq_stripe];
 567
 568
 569         if (rc) {
 570                 lov_update_set(set, req, rc);
 571                 cfs_spin_unlock(&set->set_lock);
 572                 RETURN(rc);
 573         }
 574
 575         loi->loi_id = req->rq_oi.oi_oa->o_id;
 576         loi->loi_seq = req->rq_oi.oi_oa->o_seq;
 577         loi->loi_ost_idx = req->rq_idx;
 578         loi_init(loi);
 579
 580         if (oti && set->set_cookies)
 581                 ++oti->oti_logcookies;
 582         if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLCOOKIE)
 583                 set->set_cookie_sent++;
 584
 585         lov_update_set(set, req, rc);
 586         cfs_spin_unlock(&set->set_lock);
 587
 588         CDEBUG(D_INODE, "objid "LPX64" has subobj "LPX64"/"LPU64" at idx %d\n",
 589                lsm->lsm_object_id, loi->loi_id, loi->loi_id, req->rq_idx);
 590         RETURN(rc);
 591 }
 592
 593 static int create_done(struct obd_export *exp, struct lov_request_set *set,
 594                        struct lov_stripe_md **lsmp)
 595 {
 596         struct lov_obd *lov = &exp->exp_obd->u.lov;
 597         struct obd_trans_info *oti = set->set_oti;
 598         struct obdo *src_oa = set->set_oi->oi_oa;
 599         struct lov_request *req;
 600         struct obdo *ret_oa = NULL;
 601         int success, attrset = 0, rc = 0;
 602         ENTRY;
 603
 604         LASSERT(cfs_atomic_read(&set->set_completes));
 605
 606         /* try alloc objects on other osts if osc_create fails for
 607          * exceptions: RPC failure, ENOSPC, etc */
 608         if (set->set_count != cfs_atomic_read(&set->set_success)) {
 609                 cfs_list_for_each_entry (req, &set->set_list, rq_link) {
 610                         if (req->rq_rc == 0)
 611                                 continue;
 612
 613                         cfs_atomic_dec(&set->set_completes);
 614                         req->rq_complete = 0;
 615
 616                         rc = qos_remedy_create(set, req);
 617                         lov_update_create_set(set, req, rc);
 618                 }
 619         }
 620
 621         success = cfs_atomic_read(&set->set_success);
 622         /* no successful creates */
 623         if (success == 0)
 624                 GOTO(cleanup, rc);
 625
 626         if (set->set_count != success) {
 627                 set->set_count = success;
 628                 qos_shrink_lsm(set);
 629         }
 630
 631         OBDO_ALLOC(ret_oa);
 632         if (ret_oa == NULL)
 633                 GOTO(cleanup, rc = -ENOMEM);
 634
 635         cfs_list_for_each_entry(req, &set->set_list, rq_link) {
 636                 if (!req->rq_complete || req->rq_rc)
 637                         continue;
 638                 lov_merge_attrs(ret_oa, req->rq_oi.oi_oa,
 639                                 req->rq_oi.oi_oa->o_valid, set->set_oi->oi_md,
 640                                 req->rq_stripe, &attrset);
 641         }
 642         if (src_oa->o_valid & OBD_MD_FLSIZE &&
 643             ret_oa->o_size != src_oa->o_size) {
 644                 CERROR("original size "LPU64" isn't new object size "LPU64"\n",
 645                        src_oa->o_size, ret_oa->o_size);
 646                 LBUG();
 647         }
 648         ret_oa->o_id = src_oa->o_id;
 649         ret_oa->o_seq = src_oa->o_seq;
 650         ret_oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 651         memcpy(src_oa, ret_oa, sizeof(*src_oa));
 652         OBDO_FREE(ret_oa);
 653
 654         *lsmp = set->set_oi->oi_md;
 655         GOTO(done, rc = 0);
 656
 657 cleanup:
 658         cfs_list_for_each_entry(req, &set->set_list, rq_link) {
 659                 struct obd_export *sub_exp;
 660                 int err = 0;
 661
 662                 if (!req->rq_complete || req->rq_rc)
 663                         continue;
 664
 665                 sub_exp = lov->lov_tgts[req->rq_idx]->ltd_exp;
 666                 err = obd_destroy(sub_exp, req->rq_oi.oi_oa, NULL, oti, NULL,
 667                                   NULL);
 668                 if (err)
 669                         CERROR("Failed to uncreate objid "LPX64" subobj "
 670                                LPX64" on OST idx %d: rc = %d\n",
 671                                src_oa->o_id, req->rq_oi.oi_oa->o_id,
 672                                req->rq_idx, rc);
 673         }
 674         if (*lsmp == NULL)
 675                 obd_free_memmd(exp, &set->set_oi->oi_md);
 676 done:
 677         if (oti && set->set_cookies) {
 678                 oti->oti_logcookies = set->set_cookies;
 679                 if (!set->set_cookie_sent) {
 680                         oti_free_cookies(oti);
 681                         src_oa->o_valid &= ~OBD_MD_FLCOOKIE;
 682                 } else {
 683                         src_oa->o_valid |= OBD_MD_FLCOOKIE;
 684                 }
 685         }
 686         RETURN(rc);
 687 }
 688
 689 int lov_fini_create_set(struct lov_request_set *set,struct lov_stripe_md **lsmp)
 690 {
 691         int rc = 0;
 692         ENTRY;
 693
 694         if (set == NULL)
 695                 RETURN(0);
 696         LASSERT(set->set_exp);
 697         if (cfs_atomic_read(&set->set_completes))
 698                 rc = create_done(set->set_exp, set, lsmp);
 699
 700         lov_put_reqset(set);
 701         RETURN(rc);
 702 }
 703
 704 int cb_create_update(void *cookie, int rc)
 705 {
 706         struct obd_info *oinfo = cookie;
 707         struct lov_request *lovreq;
 708
 709         lovreq = container_of(oinfo, struct lov_request, rq_oi);
 710
 711         if (CFS_FAIL_CHECK(OBD_FAIL_MDS_OSC_CREATE_FAIL))
 712                 if (lovreq->rq_idx == cfs_fail_val)
 713                         rc = -ENOTCONN;
 714
 715         rc= lov_update_create_set(lovreq->rq_rqset, lovreq, rc);
 716         if (lov_finished_set(lovreq->rq_rqset))
 717                 lov_put_reqset(lovreq->rq_rqset);
 718         return rc;
 719 }
 720
 721 int lov_prep_create_set(struct obd_export *exp, struct obd_info *oinfo,
 722                         struct lov_stripe_md **lsmp, struct obdo *src_oa,
 723                         struct obd_trans_info *oti,
 724                         struct lov_request_set **reqset)
 725 {
 726         struct lov_request_set *set;
 727         int rc = 0;
 728         ENTRY;
 729
 730         OBD_ALLOC(set, sizeof(*set));
 731         if (set == NULL)
 732                 RETURN(-ENOMEM);
 733         lov_init_set(set);
 734
 735         set->set_exp = exp;
 736         set->set_oi = oinfo;
 737         set->set_oi->oi_md = *lsmp;
 738         set->set_oi->oi_oa = src_oa;
 739         set->set_oti = oti;
 740         lov_get_reqset(set);
 741
 742         rc = qos_prep_create(exp, set);
 743         /* qos_shrink_lsm() may have allocated a new lsm */
 744         *lsmp = oinfo->oi_md;
 745         if (rc) {
 746                 lov_fini_create_set(set, lsmp);
 747                 lov_put_reqset(set);
 748         } else {
 749                 *reqset = set;
 750         }
 751         RETURN(rc);
 752 }
 753
 754 static int common_attr_done(struct lov_request_set *set)
 755 {
 756         cfs_list_t *pos;
 757         struct lov_request *req;
 758         struct obdo *tmp_oa;
 759         int rc = 0, attrset = 0;
 760         ENTRY;
 761
 762         LASSERT(set->set_oi != NULL);
 763
 764         if (set->set_oi->oi_oa == NULL)
 765                 RETURN(0);
 766
 767         if (!cfs_atomic_read(&set->set_success))
 768                 RETURN(-EIO);
 769
 770         OBDO_ALLOC(tmp_oa);
 771         if (tmp_oa == NULL)
 772                 GOTO(out, rc = -ENOMEM);
 773
 774         cfs_list_for_each (pos, &set->set_list) {
 775                 req = cfs_list_entry(pos, struct lov_request, rq_link);
 776
 777                 if (!req->rq_complete || req->rq_rc)
 778                         continue;
 779                 if (req->rq_oi.oi_oa->o_valid == 0)   /* inactive stripe */
 780                         continue;
 781                 lov_merge_attrs(tmp_oa, req->rq_oi.oi_oa,
 782                                 req->rq_oi.oi_oa->o_valid,
 783                                 set->set_oi->oi_md, req->rq_stripe, &attrset);
 784         }
 785         if (!attrset) {
 786                 CERROR("No stripes had valid attrs\n");
 787                 rc = -EIO;
 788         }
 789         if ((set->set_oi->oi_oa->o_valid & OBD_MD_FLEPOCH) &&
 790             (set->set_oi->oi_md->lsm_stripe_count != attrset)) {
 791                 /* When we take attributes of some epoch, we require all the
 792                  * ost to be active. */
 793                 CERROR("Not all the stripes had valid attrs\n");
 794                 GOTO(out, rc = -EIO);
 795         }
 796
 797         tmp_oa->o_id = set->set_oi->oi_oa->o_id;
 798         memcpy(set->set_oi->oi_oa, tmp_oa, sizeof(*set->set_oi->oi_oa));
 799 out:
 800         if (tmp_oa)
 801                 OBDO_FREE(tmp_oa);
 802         RETURN(rc);
 803
 804 }
 805
 806 static int brw_done(struct lov_request_set *set)
 807 {
 808         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 809         struct lov_oinfo     *loi = NULL;
 810         cfs_list_t *pos;
 811         struct lov_request *req;
 812         ENTRY;
 813
 814         cfs_list_for_each (pos, &set->set_list) {
 815                 req = cfs_list_entry(pos, struct lov_request, rq_link);
 816
 817                 if (!req->rq_complete || req->rq_rc)
 818                         continue;
 819
 820                 loi = lsm->lsm_oinfo[req->rq_stripe];
 821
 822                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLBLOCKS)
 823                         loi->loi_lvb.lvb_blocks = req->rq_oi.oi_oa->o_blocks;
 824         }
 825
 826         RETURN(0);
 827 }
 828
 829 int lov_fini_brw_set(struct lov_request_set *set)
 830 {
 831         int rc = 0;
 832         ENTRY;
 833
 834         if (set == NULL)
 835                 RETURN(0);
 836         LASSERT(set->set_exp);
 837         if (cfs_atomic_read(&set->set_completes)) {
 838                 rc = brw_done(set);
 839                 /* FIXME update qos data here */
 840         }
 841         lov_put_reqset(set);
 842
 843         RETURN(rc);
 844 }
 845
 846 int lov_prep_brw_set(struct obd_export *exp, struct obd_info *oinfo,
 847                      obd_count oa_bufs, struct brw_page *pga,
 848                      struct obd_trans_info *oti,
 849                      struct lov_request_set **reqset)
 850 {
 851         struct {
 852                 obd_count       index;
 853                 obd_count       count;
 854                 obd_count       off;
 855         } *info = NULL;
 856         struct lov_request_set *set;
 857         struct lov_obd *lov = &exp->exp_obd->u.lov;
 858         int rc = 0, i, shift;
 859         ENTRY;
 860
 861         OBD_ALLOC(set, sizeof(*set));
 862         if (set == NULL)
 863                 RETURN(-ENOMEM);
 864         lov_init_set(set);
 865
 866         set->set_exp = exp;
 867         set->set_oti = oti;
 868         set->set_oi = oinfo;
 869         set->set_oabufs = oa_bufs;
 870         OBD_ALLOC_LARGE(set->set_pga, oa_bufs * sizeof(*set->set_pga));
 871         if (!set->set_pga)
 872                 GOTO(out, rc = -ENOMEM);
 873
 874         OBD_ALLOC_LARGE(info, sizeof(*info) * oinfo->oi_md->lsm_stripe_count);
 875         if (!info)
 876                 GOTO(out, rc = -ENOMEM);
 877
 878         /* calculate the page count for each stripe */
 879         for (i = 0; i < oa_bufs; i++) {
 880                 int stripe = lov_stripe_number(oinfo->oi_md, pga[i].off);
 881                 info[stripe].count++;
 882         }
 883
 884         /* alloc and initialize lov request */
 885         shift = 0;
 886         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++){
 887                 struct lov_oinfo *loi = NULL;
 888                 struct lov_request *req;
 889
 890                 if (info[i].count == 0)
 891                         continue;
 892
 893                 loi = oinfo->oi_md->lsm_oinfo[i];
 894                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 895                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 896                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 897                         GOTO(out, rc = -EIO);
 898                 }
 899
 900                 OBD_ALLOC(req, sizeof(*req));
 901                 if (req == NULL)
 902                         GOTO(out, rc = -ENOMEM);
 903
 904                 OBDO_ALLOC(req->rq_oi.oi_oa);
 905                 if (req->rq_oi.oi_oa == NULL) {
 906                         OBD_FREE(req, sizeof(*req));
 907                         GOTO(out, rc = -ENOMEM);
 908                 }
 909
 910                 if (oinfo->oi_oa) {
 911                         memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
 912                                sizeof(*req->rq_oi.oi_oa));
 913                 }
 914                 req->rq_oi.oi_oa->o_id = loi->loi_id;
 915                 req->rq_oi.oi_oa->o_seq = loi->loi_seq;
 916                 req->rq_oi.oi_oa->o_stripe_idx = i;
 917
 918                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 919                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 920                 if (req->rq_oi.oi_md == NULL) {
 921                         OBDO_FREE(req->rq_oi.oi_oa);
 922                         OBD_FREE(req, sizeof(*req));
 923                         GOTO(out, rc = -ENOMEM);
 924                 }
 925
 926                 req->rq_idx = loi->loi_ost_idx;
 927                 req->rq_stripe = i;
 928
 929                 /* XXX LOV STACKING */
 930                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 931                 req->rq_oi.oi_md->lsm_object_seq = loi->loi_seq;
 932                 req->rq_oabufs = info[i].count;
 933                 req->rq_pgaidx = shift;
 934                 shift += req->rq_oabufs;
 935
 936                 /* remember the index for sort brw_page array */
 937                 info[i].index = req->rq_pgaidx;
 938
 939                 req->rq_oi.oi_capa = oinfo->oi_capa;
 940
 941                 lov_set_add_req(req, set);
 942         }
 943         if (!set->set_count)
 944                 GOTO(out, rc = -EIO);
 945
 946         /* rotate & sort the brw_page array */
 947         for (i = 0; i < oa_bufs; i++) {
 948                 int stripe = lov_stripe_number(oinfo->oi_md, pga[i].off);
 949
 950                 shift = info[stripe].index + info[stripe].off;
 951                 LASSERT(shift < oa_bufs);
 952                 set->set_pga[shift] = pga[i];
 953                 lov_stripe_offset(oinfo->oi_md, pga[i].off, stripe,
 954                                   &set->set_pga[shift].off);
 955                 info[stripe].off++;
 956         }
 957 out:
 958         if (info)
 959                 OBD_FREE_LARGE(info,
 960                                sizeof(*info) * oinfo->oi_md->lsm_stripe_count);
 961
 962         if (rc == 0)
 963                 *reqset = set;
 964         else
 965                 lov_fini_brw_set(set);
 966
 967         RETURN(rc);
 968 }
 969
 970 int lov_fini_getattr_set(struct lov_request_set *set)
 971 {
 972         int rc = 0;
 973         ENTRY;
 974
 975         if (set == NULL)
 976                 RETURN(0);
 977         LASSERT(set->set_exp);
 978         if (cfs_atomic_read(&set->set_completes))
 979                 rc = common_attr_done(set);
 980
 981         lov_put_reqset(set);
 982
 983         RETURN(rc);
 984 }
 985
 986 /* The callback for osc_getattr_async that finilizes a request info when a
 987  * response is received. */
 988 static int cb_getattr_update(void *cookie, int rc)
 989 {
 990         struct obd_info *oinfo = cookie;
 991         struct lov_request *lovreq;
 992         lovreq = container_of(oinfo, struct lov_request, rq_oi);
 993         return lov_update_common_set(lovreq->rq_rqset, lovreq, rc);
 994 }
 995
 996 int lov_prep_getattr_set(struct obd_export *exp, struct obd_info *oinfo,
 997                          struct lov_request_set **reqset)
 998 {
 999         struct lov_request_set *set;
1000         struct lov_obd *lov = &exp->exp_obd->u.lov;
1001         int rc = 0, i;
1002         ENTRY;
1003
1004         OBD_ALLOC(set, sizeof(*set));
1005         if (set == NULL)
1006                 RETURN(-ENOMEM);
1007         lov_init_set(set);
1008
1009         set->set_exp = exp;
1010         set->set_oi = oinfo;
1011
1012         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1013                 struct lov_oinfo *loi;
1014                 struct lov_request *req;
1015
1016                 loi = oinfo->oi_md->lsm_oinfo[i];
1017                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1018                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1019                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1020                         if (oinfo->oi_oa->o_valid & OBD_MD_FLEPOCH)
1021                                 /* SOM requires all the OSTs to be active. */
1022                                 GOTO(out_set, rc = -EIO);
1023                         continue;
1024                 }
1025
1026                 OBD_ALLOC(req, sizeof(*req));
1027                 if (req == NULL)
1028                         GOTO(out_set, rc = -ENOMEM);
1029
1030                 req->rq_stripe = i;
1031                 req->rq_idx = loi->loi_ost_idx;
1032
1033                 OBDO_ALLOC(req->rq_oi.oi_oa);
1034                 if (req->rq_oi.oi_oa == NULL) {
1035                         OBD_FREE(req, sizeof(*req));
1036                         GOTO(out_set, rc = -ENOMEM);
1037                 }
1038                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1039                        sizeof(*req->rq_oi.oi_oa));
1040                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1041                 req->rq_oi.oi_oa->o_seq = loi->loi_seq;
1042                 req->rq_oi.oi_cb_up = cb_getattr_update;
1043                 req->rq_oi.oi_capa = oinfo->oi_capa;
1044
1045                 lov_set_add_req(req, set);
1046         }
1047         if (!set->set_count)
1048                 GOTO(out_set, rc = -EIO);
1049         *reqset = set;
1050         RETURN(rc);
1051 out_set:
1052         lov_fini_getattr_set(set);
1053         RETURN(rc);
1054 }
1055
1056 int lov_fini_destroy_set(struct lov_request_set *set)
1057 {
1058         ENTRY;
1059
1060         if (set == NULL)
1061                 RETURN(0);
1062         LASSERT(set->set_exp);
1063         if (cfs_atomic_read(&set->set_completes)) {
1064                 /* FIXME update qos data here */
1065         }
1066
1067         lov_put_reqset(set);
1068
1069         RETURN(0);
1070 }
1071
1072 int lov_prep_destroy_set(struct obd_export *exp, struct obd_info *oinfo,
1073                          struct obdo *src_oa, struct lov_stripe_md *lsm,
1074                          struct obd_trans_info *oti,
1075                          struct lov_request_set **reqset)
1076 {
1077         struct lov_request_set *set;
1078         struct lov_obd *lov = &exp->exp_obd->u.lov;
1079         int rc = 0, i;
1080         ENTRY;
1081
1082         OBD_ALLOC(set, sizeof(*set));
1083         if (set == NULL)
1084                 RETURN(-ENOMEM);
1085         lov_init_set(set);
1086
1087         set->set_exp = exp;
1088         set->set_oi = oinfo;
1089         set->set_oi->oi_md = lsm;
1090         set->set_oi->oi_oa = src_oa;
1091         set->set_oti = oti;
1092         if (oti != NULL && src_oa->o_valid & OBD_MD_FLCOOKIE)
1093                 set->set_cookies = oti->oti_logcookies;
1094
1095         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1096                 struct lov_oinfo *loi;
1097                 struct lov_request *req;
1098
1099                 loi = lsm->lsm_oinfo[i];
1100                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1101                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1102                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1103                         continue;
1104                 }
1105
1106                 OBD_ALLOC(req, sizeof(*req));
1107                 if (req == NULL)
1108                         GOTO(out_set, rc = -ENOMEM);
1109
1110                 req->rq_stripe = i;
1111                 req->rq_idx = loi->loi_ost_idx;
1112
1113                 OBDO_ALLOC(req->rq_oi.oi_oa);
1114                 if (req->rq_oi.oi_oa == NULL) {
1115                         OBD_FREE(req, sizeof(*req));
1116                         GOTO(out_set, rc = -ENOMEM);
1117                 }
1118                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1119                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1120                 req->rq_oi.oi_oa->o_seq = loi->loi_seq;
1121                 lov_set_add_req(req, set);
1122         }
1123         if (!set->set_count)
1124                 GOTO(out_set, rc = -EIO);
1125         *reqset = set;
1126         RETURN(rc);
1127 out_set:
1128         lov_fini_destroy_set(set);
1129         RETURN(rc);
1130 }
1131
1132 int lov_fini_setattr_set(struct lov_request_set *set)
1133 {
1134         int rc = 0;
1135         ENTRY;
1136
1137         if (set == NULL)
1138                 RETURN(0);
1139         LASSERT(set->set_exp);
1140         if (cfs_atomic_read(&set->set_completes)) {
1141                 rc = common_attr_done(set);
1142                 /* FIXME update qos data here */
1143         }
1144
1145         lov_put_reqset(set);
1146         RETURN(rc);
1147 }
1148
1149 int lov_update_setattr_set(struct lov_request_set *set,
1150                            struct lov_request *req, int rc)
1151 {
1152         struct lov_obd *lov = &req->rq_rqset->set_exp->exp_obd->u.lov;
1153         struct lov_stripe_md *lsm = req->rq_rqset->set_oi->oi_md;
1154         ENTRY;
1155
1156         lov_update_set(set, req, rc);
1157
1158         /* grace error on inactive ost */
1159         if (rc && !(lov->lov_tgts[req->rq_idx] &&
1160                     lov->lov_tgts[req->rq_idx]->ltd_active))
1161                 rc = 0;
1162
1163         if (rc == 0) {
1164                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLCTIME)
1165                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_ctime =
1166                                 req->rq_oi.oi_oa->o_ctime;
1167                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLMTIME)
1168                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_mtime =
1169                                 req->rq_oi.oi_oa->o_mtime;
1170                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLATIME)
1171                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_atime =
1172                                 req->rq_oi.oi_oa->o_atime;
1173         }
1174
1175         RETURN(rc);
1176 }
1177
1178 /* The callback for osc_setattr_async that finilizes a request info when a
1179  * response is received. */
1180 static int cb_setattr_update(void *cookie, int rc)
1181 {
1182         struct obd_info *oinfo = cookie;
1183         struct lov_request *lovreq;
1184         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1185         return lov_update_setattr_set(lovreq->rq_rqset, lovreq, rc);
1186 }
1187
1188 int lov_prep_setattr_set(struct obd_export *exp, struct obd_info *oinfo,
1189                          struct obd_trans_info *oti,
1190                          struct lov_request_set **reqset)
1191 {
1192         struct lov_request_set *set;
1193         struct lov_obd *lov = &exp->exp_obd->u.lov;
1194         int rc = 0, i;
1195         ENTRY;
1196
1197         OBD_ALLOC(set, sizeof(*set));
1198         if (set == NULL)
1199                 RETURN(-ENOMEM);
1200         lov_init_set(set);
1201
1202         set->set_exp = exp;
1203         set->set_oti = oti;
1204         set->set_oi = oinfo;
1205         if (oti != NULL && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
1206                 set->set_cookies = oti->oti_logcookies;
1207
1208         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1209                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1210                 struct lov_request *req;
1211
1212                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1213                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1214                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1215                         continue;
1216                 }
1217
1218                 OBD_ALLOC(req, sizeof(*req));
1219                 if (req == NULL)
1220                         GOTO(out_set, rc = -ENOMEM);
1221                 req->rq_stripe = i;
1222                 req->rq_idx = loi->loi_ost_idx;
1223
1224                 OBDO_ALLOC(req->rq_oi.oi_oa);
1225                 if (req->rq_oi.oi_oa == NULL) {
1226                         OBD_FREE(req, sizeof(*req));
1227                         GOTO(out_set, rc = -ENOMEM);
1228                 }
1229                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1230                        sizeof(*req->rq_oi.oi_oa));
1231                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1232                 req->rq_oi.oi_oa->o_seq= loi->loi_seq;
1233                 req->rq_oi.oi_oa->o_stripe_idx = i;
1234                 req->rq_oi.oi_cb_up = cb_setattr_update;
1235                 req->rq_oi.oi_capa = oinfo->oi_capa;
1236
1237                 if (oinfo->oi_oa->o_valid & OBD_MD_FLSIZE) {
1238                         int off = lov_stripe_offset(oinfo->oi_md,
1239                                                     oinfo->oi_oa->o_size, i,
1240                                                     &req->rq_oi.oi_oa->o_size);
1241
1242                         if (off < 0 && req->rq_oi.oi_oa->o_size)
1243                                 req->rq_oi.oi_oa->o_size--;
1244
1245                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1246                                i, req->rq_oi.oi_oa->o_size,
1247                                oinfo->oi_oa->o_size);
1248                 }
1249                 lov_set_add_req(req, set);
1250         }
1251         if (!set->set_count)
1252                 GOTO(out_set, rc = -EIO);
1253         *reqset = set;
1254         RETURN(rc);
1255 out_set:
1256         lov_fini_setattr_set(set);
1257         RETURN(rc);
1258 }
1259
1260 int lov_fini_punch_set(struct lov_request_set *set)
1261 {
1262         int rc = 0;
1263         ENTRY;
1264
1265         if (set == NULL)
1266                 RETURN(0);
1267         LASSERT(set->set_exp);
1268         if (cfs_atomic_read(&set->set_completes)) {
1269                 rc = -EIO;
1270                 /* FIXME update qos data here */
1271                 if (cfs_atomic_read(&set->set_success))
1272                         rc = common_attr_done(set);
1273         }
1274
1275         lov_put_reqset(set);
1276
1277         RETURN(rc);
1278 }
1279
1280 int lov_update_punch_set(struct lov_request_set *set,
1281                          struct lov_request *req, int rc)
1282 {
1283         struct lov_obd *lov = &req->rq_rqset->set_exp->exp_obd->u.lov;
1284         struct lov_stripe_md *lsm = req->rq_rqset->set_oi->oi_md;
1285         ENTRY;
1286
1287         lov_update_set(set, req, rc);
1288
1289         /* grace error on inactive ost */
1290         if (rc && !lov->lov_tgts[req->rq_idx]->ltd_active)
1291                 rc = 0;
1292
1293         if (rc == 0) {
1294                 lov_stripe_lock(lsm);
1295                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLBLOCKS) {
1296                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_blocks =
1297                                 req->rq_oi.oi_oa->o_blocks;
1298                 }
1299
1300                 /* Do we need to update lvb_size here? It needn't because
1301                  * it have been done in ll_truncate(). -jay */
1302                 lov_stripe_unlock(lsm);
1303         }
1304
1305         RETURN(rc);
1306 }
1307
1308 /* The callback for osc_punch that finilizes a request info when a response
1309  * is received. */
1310 static int cb_update_punch(void *cookie, int rc)
1311 {
1312         struct obd_info *oinfo = cookie;
1313         struct lov_request *lovreq;
1314         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1315         return lov_update_punch_set(lovreq->rq_rqset, lovreq, rc);
1316 }
1317
1318 int lov_prep_punch_set(struct obd_export *exp, struct obd_info *oinfo,
1319                        struct obd_trans_info *oti,
1320                        struct lov_request_set **reqset)
1321 {
1322         struct lov_request_set *set;
1323         struct lov_obd *lov = &exp->exp_obd->u.lov;
1324         int rc = 0, i;
1325         ENTRY;
1326
1327         OBD_ALLOC(set, sizeof(*set));
1328         if (set == NULL)
1329                 RETURN(-ENOMEM);
1330         lov_init_set(set);
1331
1332         set->set_oi = oinfo;
1333         set->set_exp = exp;
1334
1335         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1336                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1337                 struct lov_request *req;
1338                 obd_off rs, re;
1339
1340                 if (!lov_stripe_intersects(oinfo->oi_md, i,
1341                                            oinfo->oi_policy.l_extent.start,
1342                                            oinfo->oi_policy.l_extent.end,
1343                                            &rs, &re))
1344                         continue;
1345
1346                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1347                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1348                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1349                         GOTO(out_set, rc = -EIO);
1350                 }
1351
1352                 OBD_ALLOC(req, sizeof(*req));
1353                 if (req == NULL)
1354                         GOTO(out_set, rc = -ENOMEM);
1355                 req->rq_stripe = i;
1356                 req->rq_idx = loi->loi_ost_idx;
1357
1358                 OBDO_ALLOC(req->rq_oi.oi_oa);
1359                 if (req->rq_oi.oi_oa == NULL) {
1360                         OBD_FREE(req, sizeof(*req));
1361                         GOTO(out_set, rc = -ENOMEM);
1362                 }
1363                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1364                        sizeof(*req->rq_oi.oi_oa));
1365                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1366                 req->rq_oi.oi_oa->o_seq = loi->loi_seq;
1367                 req->rq_oi.oi_oa->o_valid |= OBD_MD_FLGROUP;
1368
1369                 req->rq_oi.oi_oa->o_stripe_idx = i;
1370                 req->rq_oi.oi_cb_up = cb_update_punch;
1371
1372                 req->rq_oi.oi_policy.l_extent.start = rs;
1373                 req->rq_oi.oi_policy.l_extent.end = re;
1374                 req->rq_oi.oi_policy.l_extent.gid = -1;
1375
1376                 req->rq_oi.oi_capa = oinfo->oi_capa;
1377
1378                 lov_set_add_req(req, set);
1379         }
1380         if (!set->set_count)
1381                 GOTO(out_set, rc = -EIO);
1382         *reqset = set;
1383         RETURN(rc);
1384 out_set:
1385         lov_fini_punch_set(set);
1386         RETURN(rc);
1387 }
1388
1389 int lov_fini_sync_set(struct lov_request_set *set)
1390 {
1391         int rc = 0;
1392         ENTRY;
1393
1394         if (set == NULL)
1395                 RETURN(0);
1396         LASSERT(set->set_exp);
1397         if (cfs_atomic_read(&set->set_completes)) {
1398                 if (!cfs_atomic_read(&set->set_success))
1399                         rc = -EIO;
1400                 /* FIXME update qos data here */
1401         }
1402
1403         lov_put_reqset(set);
1404
1405         RETURN(rc);
1406 }
1407
1408 /* The callback for osc_sync that finilizes a request info when a
1409  * response is recieved. */
1410 static int cb_sync_update(void *cookie, int rc)
1411 {
1412         struct obd_info *oinfo = cookie;
1413         struct lov_request *lovreq;
1414
1415         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1416         return lov_update_common_set(lovreq->rq_rqset, lovreq, rc);
1417 }
1418
1419 int lov_prep_sync_set(struct obd_export *exp, struct obd_info *oinfo,
1420                       obd_off start, obd_off end,
1421                       struct lov_request_set **reqset)
1422 {
1423         struct lov_request_set *set;
1424         struct lov_obd *lov = &exp->exp_obd->u.lov;
1425         int rc = 0, i;
1426         ENTRY;
1427
1428         OBD_ALLOC_PTR(set);
1429         if (set == NULL)
1430                 RETURN(-ENOMEM);
1431         lov_init_set(set);
1432
1433         set->set_exp = exp;
1434         set->set_oi = oinfo;
1435
1436         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1437                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1438                 struct lov_request *req;
1439                 obd_off rs, re;
1440
1441                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1442                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1443                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1444                         continue;
1445                 }
1446
1447                 if (!lov_stripe_intersects(oinfo->oi_md, i, start, end, &rs,
1448                                            &re))
1449                         continue;
1450
1451                 OBD_ALLOC_PTR(req);
1452                 if (req == NULL)
1453                         GOTO(out_set, rc = -ENOMEM);
1454                 req->rq_stripe = i;
1455                 req->rq_idx = loi->loi_ost_idx;
1456
1457                 OBDO_ALLOC(req->rq_oi.oi_oa);
1458                 if (req->rq_oi.oi_oa == NULL) {
1459                         OBD_FREE(req, sizeof(*req));
1460                         GOTO(out_set, rc = -ENOMEM);
1461                 }
1462                 *req->rq_oi.oi_oa = *oinfo->oi_oa;
1463                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1464                 req->rq_oi.oi_oa->o_seq = loi->loi_seq;
1465                 req->rq_oi.oi_oa->o_stripe_idx = i;
1466
1467                 req->rq_oi.oi_policy.l_extent.start = rs;
1468                 req->rq_oi.oi_policy.l_extent.end = re;
1469                 req->rq_oi.oi_policy.l_extent.gid = -1;
1470                 req->rq_oi.oi_cb_up = cb_sync_update;
1471
1472                 lov_set_add_req(req, set);
1473         }
1474         if (!set->set_count)
1475                 GOTO(out_set, rc = -EIO);
1476         *reqset = set;
1477         RETURN(rc);
1478 out_set:
1479         lov_fini_sync_set(set);
1480         RETURN(rc);
1481 }
1482
1483 #define LOV_U64_MAX ((__u64)~0ULL)
1484 #define LOV_SUM_MAX(tot, add)                                           \
1485         do {                                                            \
1486                 if ((tot) + (add) < (tot))                              \
1487                         (tot) = LOV_U64_MAX;                            \
1488                 else                                                    \
1489                         (tot) += (add);                                 \
1490         } while(0)
1491
1492 int lov_fini_statfs(struct obd_device *obd, struct obd_statfs *osfs,int success)
1493 {
1494         ENTRY;
1495
1496         if (success) {
1497                 __u32 expected_stripes = lov_get_stripecnt(&obd->u.lov,
1498                                                            LOV_MAGIC, 0);
1499                 if (osfs->os_files != LOV_U64_MAX)
1500                         do_div(osfs->os_files, expected_stripes);
1501                 if (osfs->os_ffree != LOV_U64_MAX)
1502                         do_div(osfs->os_ffree, expected_stripes);
1503
1504                 cfs_spin_lock(&obd->obd_osfs_lock);
1505                 memcpy(&obd->obd_osfs, osfs, sizeof(*osfs));
1506                 obd->obd_osfs_age = cfs_time_current_64();
1507                 cfs_spin_unlock(&obd->obd_osfs_lock);
1508                 RETURN(0);
1509         }
1510
1511         RETURN(-EIO);
1512 }
1513
1514 int lov_fini_statfs_set(struct lov_request_set *set)
1515 {
1516         int rc = 0;
1517         ENTRY;
1518
1519         if (set == NULL)
1520                 RETURN(0);
1521
1522         if (cfs_atomic_read(&set->set_completes)) {
1523                 rc = lov_fini_statfs(set->set_obd, set->set_oi->oi_osfs,
1524                                      cfs_atomic_read(&set->set_success));
1525         }
1526         lov_put_reqset(set);
1527         RETURN(rc);
1528 }
1529
1530 void lov_update_statfs(struct obd_statfs *osfs, struct obd_statfs *lov_sfs,
1531                        int success)
1532 {
1533         int shift = 0, quit = 0;
1534         __u64 tmp;
1535
1536         if (success == 0) {
1537                 memcpy(osfs, lov_sfs, sizeof(*lov_sfs));
1538         } else {
1539                 if (osfs->os_bsize != lov_sfs->os_bsize) {
1540                         /* assume all block sizes are always powers of 2 */
1541                         /* get the bits difference */
1542                         tmp = osfs->os_bsize | lov_sfs->os_bsize;
1543                         for (shift = 0; shift <= 64; ++shift) {
1544                                 if (tmp & 1) {
1545                                         if (quit)
1546                                                 break;
1547                                         else
1548                                                 quit = 1;
1549                                         shift = 0;
1550                                 }
1551                                 tmp >>= 1;
1552                         }
1553                 }
1554
1555                 if (osfs->os_bsize < lov_sfs->os_bsize) {
1556                         osfs->os_bsize = lov_sfs->os_bsize;
1557
1558                         osfs->os_bfree  >>= shift;
1559                         osfs->os_bavail >>= shift;
1560                         osfs->os_blocks >>= shift;
1561                 } else if (shift != 0) {
1562                         lov_sfs->os_bfree  >>= shift;
1563                         lov_sfs->os_bavail >>= shift;
1564                         lov_sfs->os_blocks >>= shift;
1565                 }
1566 #ifdef MIN_DF
1567                 /* Sandia requested that df (and so, statfs) only
1568                    returned minimal available space on
1569                    a single OST, so people would be able to
1570                    write this much data guaranteed. */
1571                 if (osfs->os_bavail > lov_sfs->os_bavail) {
1572                         /* Presumably if new bavail is smaller,
1573                            new bfree is bigger as well */
1574                         osfs->os_bfree = lov_sfs->os_bfree;
1575                         osfs->os_bavail = lov_sfs->os_bavail;
1576                 }
1577 #else
1578                 osfs->os_bfree += lov_sfs->os_bfree;
1579                 osfs->os_bavail += lov_sfs->os_bavail;
1580 #endif
1581                 osfs->os_blocks += lov_sfs->os_blocks;
1582                 /* XXX not sure about this one - depends on policy.
1583                  *   - could be minimum if we always stripe on all OBDs
1584                  *     (but that would be wrong for any other policy,
1585                  *     if one of the OBDs has no more objects left)
1586                  *   - could be sum if we stripe whole objects
1587                  *   - could be average, just to give a nice number
1588                  *
1589                  * To give a "reasonable" (if not wholly accurate)
1590                  * number, we divide the total number of free objects
1591                  * by expected stripe count (watch out for overflow).
1592                  */
1593                 LOV_SUM_MAX(osfs->os_files, lov_sfs->os_files);
1594                 LOV_SUM_MAX(osfs->os_ffree, lov_sfs->os_ffree);
1595         }
1596 }
1597
1598 /* The callback for osc_statfs_async that finilizes a request info when a
1599  * response is received. */
1600 static int cb_statfs_update(void *cookie, int rc)
1601 {
1602         struct obd_info *oinfo = cookie;
1603         struct lov_request *lovreq;
1604         struct lov_request_set *set;
1605         struct obd_statfs *osfs, *lov_sfs;
1606         struct lov_obd *lov;
1607         struct lov_tgt_desc *tgt;
1608         struct obd_device *lovobd, *tgtobd;
1609         int success;
1610         ENTRY;
1611
1612         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1613         set = lovreq->rq_rqset;
1614         lovobd = set->set_obd;
1615         lov = &lovobd->u.lov;
1616         osfs = set->set_oi->oi_osfs;
1617         lov_sfs = oinfo->oi_osfs;
1618         success = cfs_atomic_read(&set->set_success);
1619         /* XXX: the same is done in lov_update_common_set, however
1620            lovset->set_exp is not initialized. */
1621         lov_update_set(set, lovreq, rc);
1622         if (rc)
1623                 GOTO(out, rc);
1624
1625         obd_getref(lovobd);
1626         tgt = lov->lov_tgts[lovreq->rq_idx];
1627         if (!tgt || !tgt->ltd_active)
1628                 GOTO(out_update, rc);
1629
1630         tgtobd = class_exp2obd(tgt->ltd_exp);
1631         cfs_spin_lock(&tgtobd->obd_osfs_lock);
1632         memcpy(&tgtobd->obd_osfs, lov_sfs, sizeof(*lov_sfs));
1633         if ((oinfo->oi_flags & OBD_STATFS_FROM_CACHE) == 0)
1634                 tgtobd->obd_osfs_age = cfs_time_current_64();
1635         cfs_spin_unlock(&tgtobd->obd_osfs_lock);
1636
1637 out_update:
1638         lov_update_statfs(osfs, lov_sfs, success);
1639         qos_update(lov);
1640         obd_putref(lovobd);
1641
1642 out:
1643         if (set->set_oi->oi_flags & OBD_STATFS_PTLRPCD &&
1644             lov_finished_set(set)) {
1645                 lov_statfs_interpret(NULL, set, set->set_count !=
1646                                      cfs_atomic_read(&set->set_success));
1647                 if (lov->lov_qos.lq_statfs_in_progress)
1648                         qos_statfs_done(lov);
1649         }
1650
1651         RETURN(0);
1652 }
1653
1654 int lov_prep_statfs_set(struct obd_device *obd, struct obd_info *oinfo,
1655                         struct lov_request_set **reqset)
1656 {
1657         struct lov_request_set *set;
1658         struct lov_obd *lov = &obd->u.lov;
1659         int rc = 0, i;
1660         ENTRY;
1661
1662         OBD_ALLOC(set, sizeof(*set));
1663         if (set == NULL)
1664                 RETURN(-ENOMEM);
1665         lov_init_set(set);
1666
1667         set->set_obd = obd;
1668         set->set_oi = oinfo;
1669
1670         /* We only get block data from the OBD */
1671         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1672                 struct lov_request *req;
1673
1674                 if (!lov->lov_tgts[i] || (!lov->lov_tgts[i]->ltd_active
1675                                           && (oinfo->oi_flags & OBD_STATFS_NODELAY))) {
1676                         CDEBUG(D_HA, "lov idx %d inactive\n", i);
1677                         continue;
1678                 }
1679
1680                 /* skip targets that have been explicitely disabled by the
1681                  * administrator */
1682                 if (!lov->lov_tgts[i]->ltd_exp) {
1683                         CDEBUG(D_HA, "lov idx %d administratively disabled\n", i);
1684                         continue;
1685                 }
1686
1687                 OBD_ALLOC(req, sizeof(*req));
1688                 if (req == NULL)
1689                         GOTO(out_set, rc = -ENOMEM);
1690
1691                 OBD_ALLOC(req->rq_oi.oi_osfs, sizeof(*req->rq_oi.oi_osfs));
1692                 if (req->rq_oi.oi_osfs == NULL) {
1693                         OBD_FREE(req, sizeof(*req));
1694                         GOTO(out_set, rc = -ENOMEM);
1695                 }
1696
1697                 req->rq_idx = i;
1698                 req->rq_oi.oi_cb_up = cb_statfs_update;
1699                 req->rq_oi.oi_flags = oinfo->oi_flags;
1700
1701                 lov_set_add_req(req, set);
1702         }
1703         if (!set->set_count)
1704                 GOTO(out_set, rc = -EIO);
1705         *reqset = set;
1706         RETURN(rc);
1707 out_set:
1708         lov_fini_statfs_set(set);
1709         RETURN(rc);
1710 }