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 #ifdef __KERNEL__
  40 #include <libcfs/libcfs.h>
  41 #else
  42 #include <liblustre.h>
  43 #endif
  44
  45 #include <obd_class.h>
  46 #include <obd_lov.h>
  47 #include <lustre/lustre_idl.h>
  48
  49 #include "lov_internal.h"
  50
  51 static void lov_init_set(struct lov_request_set *set)
  52 {
  53         set->set_count = 0;
  54         cfs_atomic_set(&set->set_completes, 0);
  55         cfs_atomic_set(&set->set_success, 0);
  56         cfs_atomic_set(&set->set_finish_checked, 0);
  57         set->set_cookies = 0;
  58         CFS_INIT_LIST_HEAD(&set->set_list);
  59         cfs_atomic_set(&set->set_refcount, 1);
  60         cfs_waitq_init(&set->set_waitq);
  61         spin_lock_init(&set->set_lock);
  62 }
  63
  64 void lov_finish_set(struct lov_request_set *set)
  65 {
  66         cfs_list_t *pos, *n;
  67         ENTRY;
  68
  69         LASSERT(set);
  70         cfs_list_for_each_safe(pos, n, &set->set_list) {
  71                 struct lov_request *req = cfs_list_entry(pos,
  72                                                          struct lov_request,
  73                                                          rq_link);
  74                 cfs_list_del_init(&req->rq_link);
  75
  76                 if (req->rq_oi.oi_oa)
  77                         OBDO_FREE(req->rq_oi.oi_oa);
  78                 if (req->rq_oi.oi_md)
  79                         OBD_FREE_LARGE(req->rq_oi.oi_md, req->rq_buflen);
  80                 if (req->rq_oi.oi_osfs)
  81                         OBD_FREE(req->rq_oi.oi_osfs,
  82                                  sizeof(*req->rq_oi.oi_osfs));
  83                 OBD_FREE(req, sizeof(*req));
  84         }
  85
  86         if (set->set_pga) {
  87                 int len = set->set_oabufs * sizeof(*set->set_pga);
  88                 OBD_FREE_LARGE(set->set_pga, len);
  89         }
  90         if (set->set_lockh)
  91                 lov_llh_put(set->set_lockh);
  92
  93         OBD_FREE(set, sizeof(*set));
  94         EXIT;
  95 }
  96
  97 int lov_set_finished(struct lov_request_set *set, int idempotent)
  98 {
  99         int completes = cfs_atomic_read(&set->set_completes);
 100
 101         CDEBUG(D_INFO, "check set %d/%d\n", completes, set->set_count);
 102
 103         if (completes == set->set_count) {
 104                 if (idempotent)
 105                         return 1;
 106                 if (cfs_atomic_inc_return(&set->set_finish_checked) == 1)
 107                         return 1;
 108         }
 109         return 0;
 110 }
 111
 112 void lov_update_set(struct lov_request_set *set,
 113                     struct lov_request *req, int rc)
 114 {
 115         req->rq_complete = 1;
 116         req->rq_rc = rc;
 117
 118         cfs_atomic_inc(&set->set_completes);
 119         if (rc == 0)
 120                 cfs_atomic_inc(&set->set_success);
 121
 122         cfs_waitq_signal(&set->set_waitq);
 123 }
 124
 125 int lov_update_common_set(struct lov_request_set *set,
 126                           struct lov_request *req, int rc)
 127 {
 128         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 129         ENTRY;
 130
 131         lov_update_set(set, req, rc);
 132
 133         /* grace error on inactive ost */
 134         if (rc && !(lov->lov_tgts[req->rq_idx] &&
 135                     lov->lov_tgts[req->rq_idx]->ltd_active))
 136                 rc = 0;
 137
 138         /* FIXME in raid1 regime, should return 0 */
 139         RETURN(rc);
 140 }
 141
 142 void lov_set_add_req(struct lov_request *req, struct lov_request_set *set)
 143 {
 144         cfs_list_add_tail(&req->rq_link, &set->set_list);
 145         set->set_count++;
 146         req->rq_rqset = set;
 147 }
 148
 149 extern void osc_update_enqueue(struct lustre_handle *lov_lockhp,
 150                                struct lov_oinfo *loi, int flags,
 151                                struct ost_lvb *lvb, __u32 mode, int rc);
 152
 153 static int lov_update_enqueue_lov(struct obd_export *exp,
 154                                   struct lustre_handle *lov_lockhp,
 155                                   struct lov_oinfo *loi, int flags, int idx,
 156                                   struct ost_id *oi, int rc)
 157 {
 158         struct lov_obd *lov = &exp->exp_obd->u.lov;
 159
 160         if (rc != ELDLM_OK &&
 161             !(rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT))) {
 162                 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
 163                 if (lov->lov_tgts[idx] && lov->lov_tgts[idx]->ltd_active) {
 164                         /* -EUSERS used by OST to report file contention */
 165                         if (rc != -EINTR && rc != -EUSERS)
 166                                 CERROR("%s: enqueue objid "DOSTID" subobj"
 167                                        DOSTID" on OST idx %d: rc %d\n",
 168                                        exp->exp_obd->obd_name,
 169                                        POSTID(oi), POSTID(&loi->loi_oi),
 170                                        loi->loi_ost_idx, rc);
 171                 } else
 172                         rc = ELDLM_OK;
 173         }
 174         return rc;
 175 }
 176
 177 int lov_update_enqueue_set(struct lov_request *req, __u32 mode, int rc)
 178 {
 179         struct lov_request_set *set = req->rq_rqset;
 180         struct lustre_handle *lov_lockhp;
 181         struct obd_info *oi = set->set_oi;
 182         struct lov_oinfo *loi;
 183         ENTRY;
 184
 185         LASSERT(oi != NULL);
 186
 187         lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
 188         loi = oi->oi_md->lsm_oinfo[req->rq_stripe];
 189
 190         /* XXX LOV STACKING: OSC gets a copy, created in lov_prep_enqueue_set
 191          * and that copy can be arbitrarily out of date.
 192          *
 193          * The LOV API is due for a serious rewriting anyways, and this
 194          * can be addressed then. */
 195
 196         lov_stripe_lock(oi->oi_md);
 197         osc_update_enqueue(lov_lockhp, loi, oi->oi_flags,
 198                            &req->rq_oi.oi_md->lsm_oinfo[0]->loi_lvb, mode, rc);
 199         if (rc == ELDLM_LOCK_ABORTED && (oi->oi_flags & LDLM_FL_HAS_INTENT))
 200                 memset(lov_lockhp, 0, sizeof *lov_lockhp);
 201         rc = lov_update_enqueue_lov(set->set_exp, lov_lockhp, loi, oi->oi_flags,
 202                                     req->rq_idx, &oi->oi_md->lsm_oi, rc);
 203         lov_stripe_unlock(oi->oi_md);
 204         lov_update_set(set, req, rc);
 205         RETURN(rc);
 206 }
 207
 208 /* The callback for osc_enqueue that updates lov info for every OSC request. */
 209 static int cb_update_enqueue(void *cookie, int rc)
 210 {
 211         struct obd_info *oinfo = cookie;
 212         struct ldlm_enqueue_info *einfo;
 213         struct lov_request *lovreq;
 214
 215         lovreq = container_of(oinfo, struct lov_request, rq_oi);
 216         einfo = lovreq->rq_rqset->set_ei;
 217         return lov_update_enqueue_set(lovreq, einfo->ei_mode, rc);
 218 }
 219
 220 static int enqueue_done(struct lov_request_set *set, __u32 mode)
 221 {
 222         struct lov_request *req;
 223         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 224         int completes = cfs_atomic_read(&set->set_completes);
 225         int rc = 0;
 226         ENTRY;
 227
 228         /* enqueue/match success, just return */
 229         if (completes && completes == cfs_atomic_read(&set->set_success))
 230                 RETURN(0);
 231
 232         /* cancel enqueued/matched locks */
 233         cfs_list_for_each_entry(req, &set->set_list, rq_link) {
 234                 struct lustre_handle *lov_lockhp;
 235
 236                 if (!req->rq_complete || req->rq_rc)
 237                         continue;
 238
 239                 lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
 240                 LASSERT(lov_lockhp);
 241                 if (!lustre_handle_is_used(lov_lockhp))
 242                         continue;
 243
 244                 rc = obd_cancel(lov->lov_tgts[req->rq_idx]->ltd_exp,
 245                                 req->rq_oi.oi_md, mode, lov_lockhp);
 246                 if (rc && lov->lov_tgts[req->rq_idx] &&
 247                     lov->lov_tgts[req->rq_idx]->ltd_active)
 248                         CERROR("%s: cancelling obdjid "DOSTID" on OST"
 249                                "idx %d error: rc = %d\n",
 250                                set->set_exp->exp_obd->obd_name,
 251                                POSTID(&req->rq_oi.oi_md->lsm_oi),
 252                                req->rq_idx, rc);
 253         }
 254         if (set->set_lockh)
 255                 lov_llh_put(set->set_lockh);
 256         RETURN(rc);
 257 }
 258
 259 int lov_fini_enqueue_set(struct lov_request_set *set, __u32 mode, int rc,
 260                          struct ptlrpc_request_set *rqset)
 261 {
 262         int ret = 0;
 263         ENTRY;
 264
 265         if (set == NULL)
 266                 RETURN(0);
 267         LASSERT(set->set_exp);
 268         /* Do enqueue_done only for sync requests and if any request
 269          * succeeded. */
 270         if (!rqset) {
 271                 if (rc)
 272                         cfs_atomic_set(&set->set_completes, 0);
 273                 ret = enqueue_done(set, mode);
 274         } else if (set->set_lockh)
 275                 lov_llh_put(set->set_lockh);
 276
 277         lov_put_reqset(set);
 278
 279         RETURN(rc ? rc : ret);
 280 }
 281
 282 static void lov_llh_addref(void *llhp)
 283 {
 284         struct lov_lock_handles *llh = llhp;
 285
 286         cfs_atomic_inc(&llh->llh_refcount);
 287         CDEBUG(D_INFO, "GETting llh %p : new refcount %d\n", llh,
 288                cfs_atomic_read(&llh->llh_refcount));
 289 }
 290
 291 static struct portals_handle_ops lov_handle_ops = {
 292         .hop_addref = lov_llh_addref,
 293         .hop_free   = NULL,
 294 };
 295
 296 static struct lov_lock_handles *lov_llh_new(struct lov_stripe_md *lsm)
 297 {
 298         struct lov_lock_handles *llh;
 299
 300         OBD_ALLOC(llh, sizeof *llh +
 301                   sizeof(*llh->llh_handles) * lsm->lsm_stripe_count);
 302         if (llh == NULL)
 303                 return NULL;
 304
 305         cfs_atomic_set(&llh->llh_refcount, 2);
 306         llh->llh_stripe_count = lsm->lsm_stripe_count;
 307         CFS_INIT_LIST_HEAD(&llh->llh_handle.h_link);
 308         class_handle_hash(&llh->llh_handle, &lov_handle_ops);
 309
 310         return llh;
 311 }
 312
 313 int lov_prep_enqueue_set(struct obd_export *exp, struct obd_info *oinfo,
 314                          struct ldlm_enqueue_info *einfo,
 315                          struct lov_request_set **reqset)
 316 {
 317         struct lov_obd *lov = &exp->exp_obd->u.lov;
 318         struct lov_request_set *set;
 319         int i, rc = 0;
 320         ENTRY;
 321
 322         OBD_ALLOC(set, sizeof(*set));
 323         if (set == NULL)
 324                 RETURN(-ENOMEM);
 325         lov_init_set(set);
 326
 327         set->set_exp = exp;
 328         set->set_oi = oinfo;
 329         set->set_ei = einfo;
 330         set->set_lockh = lov_llh_new(oinfo->oi_md);
 331         if (set->set_lockh == NULL)
 332                 GOTO(out_set, rc = -ENOMEM);
 333         oinfo->oi_lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 334
 335         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
 336                 struct lov_oinfo *loi;
 337                 struct lov_request *req;
 338                 obd_off start, end;
 339
 340                 loi = oinfo->oi_md->lsm_oinfo[i];
 341                 if (!lov_stripe_intersects(oinfo->oi_md, i,
 342                                            oinfo->oi_policy.l_extent.start,
 343                                            oinfo->oi_policy.l_extent.end,
 344                                            &start, &end))
 345                         continue;
 346
 347                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 348                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 349                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 350                         continue;
 351                 }
 352
 353                 OBD_ALLOC(req, sizeof(*req));
 354                 if (req == NULL)
 355                         GOTO(out_set, rc = -ENOMEM);
 356
 357                 req->rq_buflen = sizeof(*req->rq_oi.oi_md) +
 358                         sizeof(struct lov_oinfo *) +
 359                         sizeof(struct lov_oinfo);
 360                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 361                 if (req->rq_oi.oi_md == NULL) {
 362                         OBD_FREE(req, sizeof(*req));
 363                         GOTO(out_set, rc = -ENOMEM);
 364                 }
 365                 req->rq_oi.oi_md->lsm_oinfo[0] =
 366                         ((void *)req->rq_oi.oi_md) + sizeof(*req->rq_oi.oi_md) +
 367                         sizeof(struct lov_oinfo *);
 368
 369                 /* Set lov request specific parameters. */
 370                 req->rq_oi.oi_lockh = set->set_lockh->llh_handles + i;
 371                 req->rq_oi.oi_cb_up = cb_update_enqueue;
 372                 req->rq_oi.oi_flags = oinfo->oi_flags;
 373
 374                 LASSERT(req->rq_oi.oi_lockh);
 375
 376                 req->rq_oi.oi_policy.l_extent.gid =
 377                         oinfo->oi_policy.l_extent.gid;
 378                 req->rq_oi.oi_policy.l_extent.start = start;
 379                 req->rq_oi.oi_policy.l_extent.end = end;
 380
 381                 req->rq_idx = loi->loi_ost_idx;
 382                 req->rq_stripe = i;
 383
 384                 /* XXX LOV STACKING: submd should be from the subobj */
 385                 req->rq_oi.oi_md->lsm_oi = loi->loi_oi;
 386                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 387                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_kms_valid =
 388                         loi->loi_kms_valid;
 389                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_kms = loi->loi_kms;
 390                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_lvb = loi->loi_lvb;
 391
 392                 lov_set_add_req(req, set);
 393         }
 394         if (!set->set_count)
 395                 GOTO(out_set, rc = -EIO);
 396         *reqset = set;
 397         RETURN(0);
 398 out_set:
 399         lov_fini_enqueue_set(set, einfo->ei_mode, rc, NULL);
 400         RETURN(rc);
 401 }
 402
 403 int lov_fini_match_set(struct lov_request_set *set, __u32 mode, int flags)
 404 {
 405         int rc = 0;
 406         ENTRY;
 407
 408         if (set == NULL)
 409                 RETURN(0);
 410         LASSERT(set->set_exp);
 411         rc = enqueue_done(set, mode);
 412         if ((set->set_count == cfs_atomic_read(&set->set_success)) &&
 413             (flags & LDLM_FL_TEST_LOCK))
 414                 lov_llh_put(set->set_lockh);
 415
 416         lov_put_reqset(set);
 417
 418         RETURN(rc);
 419 }
 420
 421 int lov_prep_match_set(struct obd_export *exp, struct obd_info *oinfo,
 422                        struct lov_stripe_md *lsm, ldlm_policy_data_t *policy,
 423                        __u32 mode, struct lustre_handle *lockh,
 424                        struct lov_request_set **reqset)
 425 {
 426         struct lov_obd *lov = &exp->exp_obd->u.lov;
 427         struct lov_request_set *set;
 428         int i, rc = 0;
 429         ENTRY;
 430
 431         OBD_ALLOC(set, sizeof(*set));
 432         if (set == NULL)
 433                 RETURN(-ENOMEM);
 434         lov_init_set(set);
 435
 436         set->set_exp = exp;
 437         set->set_oi = oinfo;
 438         set->set_oi->oi_md = lsm;
 439         set->set_lockh = lov_llh_new(lsm);
 440         if (set->set_lockh == NULL)
 441                 GOTO(out_set, rc = -ENOMEM);
 442         lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 443
 444         for (i = 0; i < lsm->lsm_stripe_count; i++){
 445                 struct lov_oinfo *loi;
 446                 struct lov_request *req;
 447                 obd_off start, end;
 448
 449                 loi = lsm->lsm_oinfo[i];
 450                 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
 451                                            policy->l_extent.end, &start, &end))
 452                         continue;
 453
 454                 /* FIXME raid1 should grace this error */
 455                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 456                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 457                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 458                         GOTO(out_set, rc = -EIO);
 459                 }
 460
 461                 OBD_ALLOC(req, sizeof(*req));
 462                 if (req == NULL)
 463                         GOTO(out_set, rc = -ENOMEM);
 464
 465                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 466                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 467                 if (req->rq_oi.oi_md == NULL) {
 468                         OBD_FREE(req, sizeof(*req));
 469                         GOTO(out_set, rc = -ENOMEM);
 470                 }
 471
 472                 req->rq_oi.oi_policy.l_extent.start = start;
 473                 req->rq_oi.oi_policy.l_extent.end = end;
 474                 req->rq_oi.oi_policy.l_extent.gid = policy->l_extent.gid;
 475
 476                 req->rq_idx = loi->loi_ost_idx;
 477                 req->rq_stripe = i;
 478
 479                 /* XXX LOV STACKING: submd should be from the subobj */
 480                 req->rq_oi.oi_md->lsm_oi = loi->loi_oi;
 481                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 482
 483                 lov_set_add_req(req, set);
 484         }
 485         if (!set->set_count)
 486                 GOTO(out_set, rc = -EIO);
 487         *reqset = set;
 488         RETURN(rc);
 489 out_set:
 490         lov_fini_match_set(set, mode, 0);
 491         RETURN(rc);
 492 }
 493
 494 int lov_fini_cancel_set(struct lov_request_set *set)
 495 {
 496         int rc = 0;
 497         ENTRY;
 498
 499         if (set == NULL)
 500                 RETURN(0);
 501
 502         LASSERT(set->set_exp);
 503         if (set->set_lockh)
 504                 lov_llh_put(set->set_lockh);
 505
 506         lov_put_reqset(set);
 507
 508         RETURN(rc);
 509 }
 510
 511 int lov_prep_cancel_set(struct obd_export *exp, struct obd_info *oinfo,
 512                         struct lov_stripe_md *lsm, __u32 mode,
 513                         struct lustre_handle *lockh,
 514                         struct lov_request_set **reqset)
 515 {
 516         struct lov_request_set *set;
 517         int i, rc = 0;
 518         ENTRY;
 519
 520         OBD_ALLOC(set, sizeof(*set));
 521         if (set == NULL)
 522                 RETURN(-ENOMEM);
 523         lov_init_set(set);
 524
 525         set->set_exp = exp;
 526         set->set_oi = oinfo;
 527         set->set_oi->oi_md = lsm;
 528         set->set_lockh = lov_handle2llh(lockh);
 529         if (set->set_lockh == NULL) {
 530                 CERROR("LOV: invalid lov lock handle %p\n", lockh);
 531                 GOTO(out_set, rc = -EINVAL);
 532         }
 533         lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 534
 535         for (i = 0; i < lsm->lsm_stripe_count; i++){
 536                 struct lov_request *req;
 537                 struct lustre_handle *lov_lockhp;
 538                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 539
 540                 lov_lockhp = set->set_lockh->llh_handles + i;
 541                 if (!lustre_handle_is_used(lov_lockhp)) {
 542                         CDEBUG(D_INFO, "lov idx %d subobj "DOSTID" no lock\n",
 543                                loi->loi_ost_idx, POSTID(&loi->loi_oi));
 544                         continue;
 545                 }
 546
 547                 OBD_ALLOC(req, sizeof(*req));
 548                 if (req == NULL)
 549                         GOTO(out_set, rc = -ENOMEM);
 550
 551                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 552                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 553                 if (req->rq_oi.oi_md == NULL) {
 554                         OBD_FREE(req, sizeof(*req));
 555                         GOTO(out_set, rc = -ENOMEM);
 556                 }
 557
 558                 req->rq_idx = loi->loi_ost_idx;
 559                 req->rq_stripe = i;
 560
 561                 /* XXX LOV STACKING: submd should be from the subobj */
 562                 req->rq_oi.oi_md->lsm_oi = loi->loi_oi;
 563                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 564
 565                 lov_set_add_req(req, set);
 566         }
 567         if (!set->set_count)
 568                 GOTO(out_set, rc = -EIO);
 569         *reqset = set;
 570         RETURN(rc);
 571 out_set:
 572         lov_fini_cancel_set(set);
 573         RETURN(rc);
 574 }
 575 static int common_attr_done(struct lov_request_set *set)
 576 {
 577         cfs_list_t *pos;
 578         struct lov_request *req;
 579         struct obdo *tmp_oa;
 580         int rc = 0, attrset = 0;
 581         ENTRY;
 582
 583         LASSERT(set->set_oi != NULL);
 584
 585         if (set->set_oi->oi_oa == NULL)
 586                 RETURN(0);
 587
 588         if (!cfs_atomic_read(&set->set_success))
 589                 RETURN(-EIO);
 590
 591         OBDO_ALLOC(tmp_oa);
 592         if (tmp_oa == NULL)
 593                 GOTO(out, rc = -ENOMEM);
 594
 595         cfs_list_for_each (pos, &set->set_list) {
 596                 req = cfs_list_entry(pos, struct lov_request, rq_link);
 597
 598                 if (!req->rq_complete || req->rq_rc)
 599                         continue;
 600                 if (req->rq_oi.oi_oa->o_valid == 0)   /* inactive stripe */
 601                         continue;
 602                 lov_merge_attrs(tmp_oa, req->rq_oi.oi_oa,
 603                                 req->rq_oi.oi_oa->o_valid,
 604                                 set->set_oi->oi_md, req->rq_stripe, &attrset);
 605         }
 606         if (!attrset) {
 607                 CERROR("No stripes had valid attrs\n");
 608                 rc = -EIO;
 609         }
 610         if ((set->set_oi->oi_oa->o_valid & OBD_MD_FLEPOCH) &&
 611             (set->set_oi->oi_md->lsm_stripe_count != attrset)) {
 612                 /* When we take attributes of some epoch, we require all the
 613                  * ost to be active. */
 614                 CERROR("Not all the stripes had valid attrs\n");
 615                 GOTO(out, rc = -EIO);
 616         }
 617
 618         tmp_oa->o_oi = set->set_oi->oi_oa->o_oi;
 619         memcpy(set->set_oi->oi_oa, tmp_oa, sizeof(*set->set_oi->oi_oa));
 620 out:
 621         if (tmp_oa)
 622                 OBDO_FREE(tmp_oa);
 623         RETURN(rc);
 624
 625 }
 626
 627 static int brw_done(struct lov_request_set *set)
 628 {
 629         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 630         struct lov_oinfo     *loi = NULL;
 631         cfs_list_t *pos;
 632         struct lov_request *req;
 633         ENTRY;
 634
 635         cfs_list_for_each (pos, &set->set_list) {
 636                 req = cfs_list_entry(pos, struct lov_request, rq_link);
 637
 638                 if (!req->rq_complete || req->rq_rc)
 639                         continue;
 640
 641                 loi = lsm->lsm_oinfo[req->rq_stripe];
 642
 643                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLBLOCKS)
 644                         loi->loi_lvb.lvb_blocks = req->rq_oi.oi_oa->o_blocks;
 645         }
 646
 647         RETURN(0);
 648 }
 649
 650 int lov_fini_brw_set(struct lov_request_set *set)
 651 {
 652         int rc = 0;
 653         ENTRY;
 654
 655         if (set == NULL)
 656                 RETURN(0);
 657         LASSERT(set->set_exp);
 658         if (cfs_atomic_read(&set->set_completes)) {
 659                 rc = brw_done(set);
 660                 /* FIXME update qos data here */
 661         }
 662         lov_put_reqset(set);
 663
 664         RETURN(rc);
 665 }
 666
 667 int lov_prep_brw_set(struct obd_export *exp, struct obd_info *oinfo,
 668                      obd_count oa_bufs, struct brw_page *pga,
 669                      struct obd_trans_info *oti,
 670                      struct lov_request_set **reqset)
 671 {
 672         struct {
 673                 obd_count       index;
 674                 obd_count       count;
 675                 obd_count       off;
 676         } *info = NULL;
 677         struct lov_request_set *set;
 678         struct lov_obd *lov = &exp->exp_obd->u.lov;
 679         int rc = 0, i, shift;
 680         ENTRY;
 681
 682         OBD_ALLOC(set, sizeof(*set));
 683         if (set == NULL)
 684                 RETURN(-ENOMEM);
 685         lov_init_set(set);
 686
 687         set->set_exp = exp;
 688         set->set_oti = oti;
 689         set->set_oi = oinfo;
 690         set->set_oabufs = oa_bufs;
 691         OBD_ALLOC_LARGE(set->set_pga, oa_bufs * sizeof(*set->set_pga));
 692         if (!set->set_pga)
 693                 GOTO(out, rc = -ENOMEM);
 694
 695         OBD_ALLOC_LARGE(info, sizeof(*info) * oinfo->oi_md->lsm_stripe_count);
 696         if (!info)
 697                 GOTO(out, rc = -ENOMEM);
 698
 699         /* calculate the page count for each stripe */
 700         for (i = 0; i < oa_bufs; i++) {
 701                 int stripe = lov_stripe_number(oinfo->oi_md, pga[i].off);
 702                 info[stripe].count++;
 703         }
 704
 705         /* alloc and initialize lov request */
 706         shift = 0;
 707         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++){
 708                 struct lov_oinfo *loi = NULL;
 709                 struct lov_request *req;
 710
 711                 if (info[i].count == 0)
 712                         continue;
 713
 714                 loi = oinfo->oi_md->lsm_oinfo[i];
 715                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 716                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 717                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 718                         GOTO(out, rc = -EIO);
 719                 }
 720
 721                 OBD_ALLOC(req, sizeof(*req));
 722                 if (req == NULL)
 723                         GOTO(out, rc = -ENOMEM);
 724
 725                 OBDO_ALLOC(req->rq_oi.oi_oa);
 726                 if (req->rq_oi.oi_oa == NULL) {
 727                         OBD_FREE(req, sizeof(*req));
 728                         GOTO(out, rc = -ENOMEM);
 729                 }
 730
 731                 if (oinfo->oi_oa) {
 732                         memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
 733                                sizeof(*req->rq_oi.oi_oa));
 734                 }
 735                 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
 736                 req->rq_oi.oi_oa->o_stripe_idx = i;
 737
 738                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 739                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 740                 if (req->rq_oi.oi_md == NULL) {
 741                         OBDO_FREE(req->rq_oi.oi_oa);
 742                         OBD_FREE(req, sizeof(*req));
 743                         GOTO(out, rc = -ENOMEM);
 744                 }
 745
 746                 req->rq_idx = loi->loi_ost_idx;
 747                 req->rq_stripe = i;
 748
 749                 /* XXX LOV STACKING */
 750                 req->rq_oi.oi_md->lsm_oi = loi->loi_oi;
 751                 req->rq_oabufs = info[i].count;
 752                 req->rq_pgaidx = shift;
 753                 shift += req->rq_oabufs;
 754
 755                 /* remember the index for sort brw_page array */
 756                 info[i].index = req->rq_pgaidx;
 757
 758                 req->rq_oi.oi_capa = oinfo->oi_capa;
 759
 760                 lov_set_add_req(req, set);
 761         }
 762         if (!set->set_count)
 763                 GOTO(out, rc = -EIO);
 764
 765         /* rotate & sort the brw_page array */
 766         for (i = 0; i < oa_bufs; i++) {
 767                 int stripe = lov_stripe_number(oinfo->oi_md, pga[i].off);
 768
 769                 shift = info[stripe].index + info[stripe].off;
 770                 LASSERT(shift < oa_bufs);
 771                 set->set_pga[shift] = pga[i];
 772                 lov_stripe_offset(oinfo->oi_md, pga[i].off, stripe,
 773                                   &set->set_pga[shift].off);
 774                 info[stripe].off++;
 775         }
 776 out:
 777         if (info)
 778                 OBD_FREE_LARGE(info,
 779                                sizeof(*info) * oinfo->oi_md->lsm_stripe_count);
 780
 781         if (rc == 0)
 782                 *reqset = set;
 783         else
 784                 lov_fini_brw_set(set);
 785
 786         RETURN(rc);
 787 }
 788
 789 int lov_fini_getattr_set(struct lov_request_set *set)
 790 {
 791         int rc = 0;
 792         ENTRY;
 793
 794         if (set == NULL)
 795                 RETURN(0);
 796         LASSERT(set->set_exp);
 797         if (cfs_atomic_read(&set->set_completes))
 798                 rc = common_attr_done(set);
 799
 800         lov_put_reqset(set);
 801
 802         RETURN(rc);
 803 }
 804
 805 /* The callback for osc_getattr_async that finilizes a request info when a
 806  * response is received. */
 807 static int cb_getattr_update(void *cookie, int rc)
 808 {
 809         struct obd_info *oinfo = cookie;
 810         struct lov_request *lovreq;
 811         lovreq = container_of(oinfo, struct lov_request, rq_oi);
 812         return lov_update_common_set(lovreq->rq_rqset, lovreq, rc);
 813 }
 814
 815 int lov_prep_getattr_set(struct obd_export *exp, struct obd_info *oinfo,
 816                          struct lov_request_set **reqset)
 817 {
 818         struct lov_request_set *set;
 819         struct lov_obd *lov = &exp->exp_obd->u.lov;
 820         int rc = 0, i;
 821         ENTRY;
 822
 823         OBD_ALLOC(set, sizeof(*set));
 824         if (set == NULL)
 825                 RETURN(-ENOMEM);
 826         lov_init_set(set);
 827
 828         set->set_exp = exp;
 829         set->set_oi = oinfo;
 830
 831         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
 832                 struct lov_oinfo *loi;
 833                 struct lov_request *req;
 834
 835                 loi = oinfo->oi_md->lsm_oinfo[i];
 836                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 837                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 838                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 839                         if (oinfo->oi_oa->o_valid & OBD_MD_FLEPOCH)
 840                                 /* SOM requires all the OSTs to be active. */
 841                                 GOTO(out_set, rc = -EIO);
 842                         continue;
 843                 }
 844
 845                 OBD_ALLOC(req, sizeof(*req));
 846                 if (req == NULL)
 847                         GOTO(out_set, rc = -ENOMEM);
 848
 849                 req->rq_stripe = i;
 850                 req->rq_idx = loi->loi_ost_idx;
 851
 852                 OBDO_ALLOC(req->rq_oi.oi_oa);
 853                 if (req->rq_oi.oi_oa == NULL) {
 854                         OBD_FREE(req, sizeof(*req));
 855                         GOTO(out_set, rc = -ENOMEM);
 856                 }
 857                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
 858                        sizeof(*req->rq_oi.oi_oa));
 859                 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
 860                 req->rq_oi.oi_cb_up = cb_getattr_update;
 861                 req->rq_oi.oi_capa = oinfo->oi_capa;
 862
 863                 lov_set_add_req(req, set);
 864         }
 865         if (!set->set_count)
 866                 GOTO(out_set, rc = -EIO);
 867         *reqset = set;
 868         RETURN(rc);
 869 out_set:
 870         lov_fini_getattr_set(set);
 871         RETURN(rc);
 872 }
 873
 874 int lov_fini_destroy_set(struct lov_request_set *set)
 875 {
 876         ENTRY;
 877
 878         if (set == NULL)
 879                 RETURN(0);
 880         LASSERT(set->set_exp);
 881         if (cfs_atomic_read(&set->set_completes)) {
 882                 /* FIXME update qos data here */
 883         }
 884
 885         lov_put_reqset(set);
 886
 887         RETURN(0);
 888 }
 889
 890 int lov_prep_destroy_set(struct obd_export *exp, struct obd_info *oinfo,
 891                          struct obdo *src_oa, struct lov_stripe_md *lsm,
 892                          struct obd_trans_info *oti,
 893                          struct lov_request_set **reqset)
 894 {
 895         struct lov_request_set *set;
 896         struct lov_obd *lov = &exp->exp_obd->u.lov;
 897         int rc = 0, i;
 898         ENTRY;
 899
 900         OBD_ALLOC(set, sizeof(*set));
 901         if (set == NULL)
 902                 RETURN(-ENOMEM);
 903         lov_init_set(set);
 904
 905         set->set_exp = exp;
 906         set->set_oi = oinfo;
 907         set->set_oi->oi_md = lsm;
 908         set->set_oi->oi_oa = src_oa;
 909         set->set_oti = oti;
 910         if (oti != NULL && src_oa->o_valid & OBD_MD_FLCOOKIE)
 911                 set->set_cookies = oti->oti_logcookies;
 912
 913         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 914                 struct lov_oinfo *loi;
 915                 struct lov_request *req;
 916
 917                 loi = lsm->lsm_oinfo[i];
 918                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 919                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 920                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 921                         continue;
 922                 }
 923
 924                 OBD_ALLOC(req, sizeof(*req));
 925                 if (req == NULL)
 926                         GOTO(out_set, rc = -ENOMEM);
 927
 928                 req->rq_stripe = i;
 929                 req->rq_idx = loi->loi_ost_idx;
 930
 931                 OBDO_ALLOC(req->rq_oi.oi_oa);
 932                 if (req->rq_oi.oi_oa == NULL) {
 933                         OBD_FREE(req, sizeof(*req));
 934                         GOTO(out_set, rc = -ENOMEM);
 935                 }
 936                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
 937                 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
 938                 lov_set_add_req(req, set);
 939         }
 940         if (!set->set_count)
 941                 GOTO(out_set, rc = -EIO);
 942         *reqset = set;
 943         RETURN(rc);
 944 out_set:
 945         lov_fini_destroy_set(set);
 946         RETURN(rc);
 947 }
 948
 949 int lov_fini_setattr_set(struct lov_request_set *set)
 950 {
 951         int rc = 0;
 952         ENTRY;
 953
 954         if (set == NULL)
 955                 RETURN(0);
 956         LASSERT(set->set_exp);
 957         if (cfs_atomic_read(&set->set_completes)) {
 958                 rc = common_attr_done(set);
 959                 /* FIXME update qos data here */
 960         }
 961
 962         lov_put_reqset(set);
 963         RETURN(rc);
 964 }
 965
 966 int lov_update_setattr_set(struct lov_request_set *set,
 967                            struct lov_request *req, int rc)
 968 {
 969         struct lov_obd *lov = &req->rq_rqset->set_exp->exp_obd->u.lov;
 970         struct lov_stripe_md *lsm = req->rq_rqset->set_oi->oi_md;
 971         ENTRY;
 972
 973         lov_update_set(set, req, rc);
 974
 975         /* grace error on inactive ost */
 976         if (rc && !(lov->lov_tgts[req->rq_idx] &&
 977                     lov->lov_tgts[req->rq_idx]->ltd_active))
 978                 rc = 0;
 979
 980         if (rc == 0) {
 981                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLCTIME)
 982                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_ctime =
 983                                 req->rq_oi.oi_oa->o_ctime;
 984                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLMTIME)
 985                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_mtime =
 986                                 req->rq_oi.oi_oa->o_mtime;
 987                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLATIME)
 988                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_atime =
 989                                 req->rq_oi.oi_oa->o_atime;
 990         }
 991
 992         RETURN(rc);
 993 }
 994
 995 /* The callback for osc_setattr_async that finilizes a request info when a
 996  * response is received. */
 997 static int cb_setattr_update(void *cookie, int rc)
 998 {
 999         struct obd_info *oinfo = cookie;
1000         struct lov_request *lovreq;
1001         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1002         return lov_update_setattr_set(lovreq->rq_rqset, lovreq, rc);
1003 }
1004
1005 int lov_prep_setattr_set(struct obd_export *exp, struct obd_info *oinfo,
1006                          struct obd_trans_info *oti,
1007                          struct lov_request_set **reqset)
1008 {
1009         struct lov_request_set *set;
1010         struct lov_obd *lov = &exp->exp_obd->u.lov;
1011         int rc = 0, i;
1012         ENTRY;
1013
1014         OBD_ALLOC(set, sizeof(*set));
1015         if (set == NULL)
1016                 RETURN(-ENOMEM);
1017         lov_init_set(set);
1018
1019         set->set_exp = exp;
1020         set->set_oti = oti;
1021         set->set_oi = oinfo;
1022         if (oti != NULL && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
1023                 set->set_cookies = oti->oti_logcookies;
1024
1025         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1026                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1027                 struct lov_request *req;
1028
1029                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1030                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1031                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1032                         continue;
1033                 }
1034
1035                 OBD_ALLOC(req, sizeof(*req));
1036                 if (req == NULL)
1037                         GOTO(out_set, rc = -ENOMEM);
1038                 req->rq_stripe = i;
1039                 req->rq_idx = loi->loi_ost_idx;
1040
1041                 OBDO_ALLOC(req->rq_oi.oi_oa);
1042                 if (req->rq_oi.oi_oa == NULL) {
1043                         OBD_FREE(req, sizeof(*req));
1044                         GOTO(out_set, rc = -ENOMEM);
1045                 }
1046                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1047                        sizeof(*req->rq_oi.oi_oa));
1048                 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
1049                 req->rq_oi.oi_oa->o_stripe_idx = i;
1050                 req->rq_oi.oi_cb_up = cb_setattr_update;
1051                 req->rq_oi.oi_capa = oinfo->oi_capa;
1052
1053                 if (oinfo->oi_oa->o_valid & OBD_MD_FLSIZE) {
1054                         int off = lov_stripe_offset(oinfo->oi_md,
1055                                                     oinfo->oi_oa->o_size, i,
1056                                                     &req->rq_oi.oi_oa->o_size);
1057
1058                         if (off < 0 && req->rq_oi.oi_oa->o_size)
1059                                 req->rq_oi.oi_oa->o_size--;
1060
1061                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1062                                i, req->rq_oi.oi_oa->o_size,
1063                                oinfo->oi_oa->o_size);
1064                 }
1065                 lov_set_add_req(req, set);
1066         }
1067         if (!set->set_count)
1068                 GOTO(out_set, rc = -EIO);
1069         *reqset = set;
1070         RETURN(rc);
1071 out_set:
1072         lov_fini_setattr_set(set);
1073         RETURN(rc);
1074 }
1075
1076 int lov_fini_punch_set(struct lov_request_set *set)
1077 {
1078         int rc = 0;
1079         ENTRY;
1080
1081         if (set == NULL)
1082                 RETURN(0);
1083         LASSERT(set->set_exp);
1084         if (cfs_atomic_read(&set->set_completes)) {
1085                 rc = -EIO;
1086                 /* FIXME update qos data here */
1087                 if (cfs_atomic_read(&set->set_success))
1088                         rc = common_attr_done(set);
1089         }
1090
1091         lov_put_reqset(set);
1092
1093         RETURN(rc);
1094 }
1095
1096 int lov_update_punch_set(struct lov_request_set *set,
1097                          struct lov_request *req, int rc)
1098 {
1099         struct lov_obd *lov = &req->rq_rqset->set_exp->exp_obd->u.lov;
1100         struct lov_stripe_md *lsm = req->rq_rqset->set_oi->oi_md;
1101         ENTRY;
1102
1103         lov_update_set(set, req, rc);
1104
1105         /* grace error on inactive ost */
1106         if (rc && !lov->lov_tgts[req->rq_idx]->ltd_active)
1107                 rc = 0;
1108
1109         if (rc == 0) {
1110                 lov_stripe_lock(lsm);
1111                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLBLOCKS) {
1112                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_blocks =
1113                                 req->rq_oi.oi_oa->o_blocks;
1114                 }
1115
1116                 lov_stripe_unlock(lsm);
1117         }
1118
1119         RETURN(rc);
1120 }
1121
1122 /* The callback for osc_punch that finilizes a request info when a response
1123  * is received. */
1124 static int cb_update_punch(void *cookie, int rc)
1125 {
1126         struct obd_info *oinfo = cookie;
1127         struct lov_request *lovreq;
1128         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1129         return lov_update_punch_set(lovreq->rq_rqset, lovreq, rc);
1130 }
1131
1132 int lov_prep_punch_set(struct obd_export *exp, struct obd_info *oinfo,
1133                        struct obd_trans_info *oti,
1134                        struct lov_request_set **reqset)
1135 {
1136         struct lov_request_set *set;
1137         struct lov_obd *lov = &exp->exp_obd->u.lov;
1138         int rc = 0, i;
1139         ENTRY;
1140
1141         OBD_ALLOC(set, sizeof(*set));
1142         if (set == NULL)
1143                 RETURN(-ENOMEM);
1144         lov_init_set(set);
1145
1146         set->set_oi = oinfo;
1147         set->set_exp = exp;
1148
1149         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1150                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1151                 struct lov_request *req;
1152                 obd_off rs, re;
1153
1154                 if (!lov_stripe_intersects(oinfo->oi_md, i,
1155                                            oinfo->oi_policy.l_extent.start,
1156                                            oinfo->oi_policy.l_extent.end,
1157                                            &rs, &re))
1158                         continue;
1159
1160                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1161                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1162                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1163                         GOTO(out_set, rc = -EIO);
1164                 }
1165
1166                 OBD_ALLOC(req, sizeof(*req));
1167                 if (req == NULL)
1168                         GOTO(out_set, rc = -ENOMEM);
1169                 req->rq_stripe = i;
1170                 req->rq_idx = loi->loi_ost_idx;
1171
1172                 OBDO_ALLOC(req->rq_oi.oi_oa);
1173                 if (req->rq_oi.oi_oa == NULL) {
1174                         OBD_FREE(req, sizeof(*req));
1175                         GOTO(out_set, rc = -ENOMEM);
1176                 }
1177                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1178                        sizeof(*req->rq_oi.oi_oa));
1179                 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
1180                 req->rq_oi.oi_oa->o_valid |= OBD_MD_FLGROUP;
1181
1182                 req->rq_oi.oi_oa->o_stripe_idx = i;
1183                 req->rq_oi.oi_cb_up = cb_update_punch;
1184
1185                 req->rq_oi.oi_policy.l_extent.start = rs;
1186                 req->rq_oi.oi_policy.l_extent.end = re;
1187                 req->rq_oi.oi_policy.l_extent.gid = -1;
1188
1189                 req->rq_oi.oi_capa = oinfo->oi_capa;
1190
1191                 lov_set_add_req(req, set);
1192         }
1193         if (!set->set_count)
1194                 GOTO(out_set, rc = -EIO);
1195         *reqset = set;
1196         RETURN(rc);
1197 out_set:
1198         lov_fini_punch_set(set);
1199         RETURN(rc);
1200 }
1201
1202 int lov_fini_sync_set(struct lov_request_set *set)
1203 {
1204         int rc = 0;
1205         ENTRY;
1206
1207         if (set == NULL)
1208                 RETURN(0);
1209         LASSERT(set->set_exp);
1210         if (cfs_atomic_read(&set->set_completes)) {
1211                 if (!cfs_atomic_read(&set->set_success))
1212                         rc = -EIO;
1213                 /* FIXME update qos data here */
1214         }
1215
1216         lov_put_reqset(set);
1217
1218         RETURN(rc);
1219 }
1220
1221 /* The callback for osc_sync that finilizes a request info when a
1222  * response is recieved. */
1223 static int cb_sync_update(void *cookie, int rc)
1224 {
1225         struct obd_info *oinfo = cookie;
1226         struct lov_request *lovreq;
1227
1228         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1229         return lov_update_common_set(lovreq->rq_rqset, lovreq, rc);
1230 }
1231
1232 int lov_prep_sync_set(struct obd_export *exp, struct obd_info *oinfo,
1233                       obd_off start, obd_off end,
1234                       struct lov_request_set **reqset)
1235 {
1236         struct lov_request_set *set;
1237         struct lov_obd *lov = &exp->exp_obd->u.lov;
1238         int rc = 0, i;
1239         ENTRY;
1240
1241         OBD_ALLOC_PTR(set);
1242         if (set == NULL)
1243                 RETURN(-ENOMEM);
1244         lov_init_set(set);
1245
1246         set->set_exp = exp;
1247         set->set_oi = oinfo;
1248
1249         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1250                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1251                 struct lov_request *req;
1252                 obd_off rs, re;
1253
1254                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1255                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1256                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1257                         continue;
1258                 }
1259
1260                 if (!lov_stripe_intersects(oinfo->oi_md, i, start, end, &rs,
1261                                            &re))
1262                         continue;
1263
1264                 OBD_ALLOC_PTR(req);
1265                 if (req == NULL)
1266                         GOTO(out_set, rc = -ENOMEM);
1267                 req->rq_stripe = i;
1268                 req->rq_idx = loi->loi_ost_idx;
1269
1270                 OBDO_ALLOC(req->rq_oi.oi_oa);
1271                 if (req->rq_oi.oi_oa == NULL) {
1272                         OBD_FREE(req, sizeof(*req));
1273                         GOTO(out_set, rc = -ENOMEM);
1274                 }
1275                 *req->rq_oi.oi_oa = *oinfo->oi_oa;
1276                 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
1277                 req->rq_oi.oi_oa->o_stripe_idx = i;
1278
1279                 req->rq_oi.oi_policy.l_extent.start = rs;
1280                 req->rq_oi.oi_policy.l_extent.end = re;
1281                 req->rq_oi.oi_policy.l_extent.gid = -1;
1282                 req->rq_oi.oi_cb_up = cb_sync_update;
1283
1284                 lov_set_add_req(req, set);
1285         }
1286         if (!set->set_count)
1287                 GOTO(out_set, rc = -EIO);
1288         *reqset = set;
1289         RETURN(rc);
1290 out_set:
1291         lov_fini_sync_set(set);
1292         RETURN(rc);
1293 }
1294
1295 #define LOV_U64_MAX ((__u64)~0ULL)
1296 #define LOV_SUM_MAX(tot, add)                                           \
1297         do {                                                            \
1298                 if ((tot) + (add) < (tot))                              \
1299                         (tot) = LOV_U64_MAX;                            \
1300                 else                                                    \
1301                         (tot) += (add);                                 \
1302         } while(0)
1303
1304 int lov_fini_statfs(struct obd_device *obd, struct obd_statfs *osfs,int success)
1305 {
1306         ENTRY;
1307
1308         if (success) {
1309                 __u32 expected_stripes = lov_get_stripecnt(&obd->u.lov,
1310                                                            LOV_MAGIC, 0);
1311                 if (osfs->os_files != LOV_U64_MAX)
1312                         lov_do_div64(osfs->os_files, expected_stripes);
1313                 if (osfs->os_ffree != LOV_U64_MAX)
1314                         lov_do_div64(osfs->os_ffree, expected_stripes);
1315
1316                 spin_lock(&obd->obd_osfs_lock);
1317                 memcpy(&obd->obd_osfs, osfs, sizeof(*osfs));
1318                 obd->obd_osfs_age = cfs_time_current_64();
1319                 spin_unlock(&obd->obd_osfs_lock);
1320                 RETURN(0);
1321         }
1322
1323         RETURN(-EIO);
1324 }
1325
1326 int lov_fini_statfs_set(struct lov_request_set *set)
1327 {
1328         int rc = 0;
1329         ENTRY;
1330
1331         if (set == NULL)
1332                 RETURN(0);
1333
1334         if (cfs_atomic_read(&set->set_completes)) {
1335                 rc = lov_fini_statfs(set->set_obd, set->set_oi->oi_osfs,
1336                                      cfs_atomic_read(&set->set_success));
1337         }
1338         lov_put_reqset(set);
1339         RETURN(rc);
1340 }
1341
1342 void lov_update_statfs(struct obd_statfs *osfs, struct obd_statfs *lov_sfs,
1343                        int success)
1344 {
1345         int shift = 0, quit = 0;
1346         __u64 tmp;
1347
1348         if (success == 0) {
1349                 memcpy(osfs, lov_sfs, sizeof(*lov_sfs));
1350         } else {
1351                 if (osfs->os_bsize != lov_sfs->os_bsize) {
1352                         /* assume all block sizes are always powers of 2 */
1353                         /* get the bits difference */
1354                         tmp = osfs->os_bsize | lov_sfs->os_bsize;
1355                         for (shift = 0; shift <= 64; ++shift) {
1356                                 if (tmp & 1) {
1357                                         if (quit)
1358                                                 break;
1359                                         else
1360                                                 quit = 1;
1361                                         shift = 0;
1362                                 }
1363                                 tmp >>= 1;
1364                         }
1365                 }
1366
1367                 if (osfs->os_bsize < lov_sfs->os_bsize) {
1368                         osfs->os_bsize = lov_sfs->os_bsize;
1369
1370                         osfs->os_bfree  >>= shift;
1371                         osfs->os_bavail >>= shift;
1372                         osfs->os_blocks >>= shift;
1373                 } else if (shift != 0) {
1374                         lov_sfs->os_bfree  >>= shift;
1375                         lov_sfs->os_bavail >>= shift;
1376                         lov_sfs->os_blocks >>= shift;
1377                 }
1378 #ifdef MIN_DF
1379                 /* Sandia requested that df (and so, statfs) only
1380                    returned minimal available space on
1381                    a single OST, so people would be able to
1382                    write this much data guaranteed. */
1383                 if (osfs->os_bavail > lov_sfs->os_bavail) {
1384                         /* Presumably if new bavail is smaller,
1385                            new bfree is bigger as well */
1386                         osfs->os_bfree = lov_sfs->os_bfree;
1387                         osfs->os_bavail = lov_sfs->os_bavail;
1388                 }
1389 #else
1390                 osfs->os_bfree += lov_sfs->os_bfree;
1391                 osfs->os_bavail += lov_sfs->os_bavail;
1392 #endif
1393                 osfs->os_blocks += lov_sfs->os_blocks;
1394                 /* XXX not sure about this one - depends on policy.
1395                  *   - could be minimum if we always stripe on all OBDs
1396                  *     (but that would be wrong for any other policy,
1397                  *     if one of the OBDs has no more objects left)
1398                  *   - could be sum if we stripe whole objects
1399                  *   - could be average, just to give a nice number
1400                  *
1401                  * To give a "reasonable" (if not wholly accurate)
1402                  * number, we divide the total number of free objects
1403                  * by expected stripe count (watch out for overflow).
1404                  */
1405                 LOV_SUM_MAX(osfs->os_files, lov_sfs->os_files);
1406                 LOV_SUM_MAX(osfs->os_ffree, lov_sfs->os_ffree);
1407         }
1408 }
1409
1410 /* The callback for osc_statfs_async that finilizes a request info when a
1411  * response is received. */
1412 static int cb_statfs_update(void *cookie, int rc)
1413 {
1414         struct obd_info *oinfo = cookie;
1415         struct lov_request *lovreq;
1416         struct lov_request_set *set;
1417         struct obd_statfs *osfs, *lov_sfs;
1418         struct lov_obd *lov;
1419         struct lov_tgt_desc *tgt;
1420         struct obd_device *lovobd, *tgtobd;
1421         int success;
1422         ENTRY;
1423
1424         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1425         set = lovreq->rq_rqset;
1426         lovobd = set->set_obd;
1427         lov = &lovobd->u.lov;
1428         osfs = set->set_oi->oi_osfs;
1429         lov_sfs = oinfo->oi_osfs;
1430         success = cfs_atomic_read(&set->set_success);
1431         /* XXX: the same is done in lov_update_common_set, however
1432            lovset->set_exp is not initialized. */
1433         lov_update_set(set, lovreq, rc);
1434         if (rc)
1435                 GOTO(out, rc);
1436
1437         obd_getref(lovobd);
1438         tgt = lov->lov_tgts[lovreq->rq_idx];
1439         if (!tgt || !tgt->ltd_active)
1440                 GOTO(out_update, rc);
1441
1442         tgtobd = class_exp2obd(tgt->ltd_exp);
1443         spin_lock(&tgtobd->obd_osfs_lock);
1444         memcpy(&tgtobd->obd_osfs, lov_sfs, sizeof(*lov_sfs));
1445         if ((oinfo->oi_flags & OBD_STATFS_FROM_CACHE) == 0)
1446                 tgtobd->obd_osfs_age = cfs_time_current_64();
1447         spin_unlock(&tgtobd->obd_osfs_lock);
1448
1449 out_update:
1450         lov_update_statfs(osfs, lov_sfs, success);
1451         obd_putref(lovobd);
1452
1453 out:
1454         if (set->set_oi->oi_flags & OBD_STATFS_PTLRPCD &&
1455             lov_set_finished(set, 0)) {
1456                 lov_statfs_interpret(NULL, set, set->set_count !=
1457                                      cfs_atomic_read(&set->set_success));
1458         }
1459
1460         RETURN(0);
1461 }
1462
1463 int lov_prep_statfs_set(struct obd_device *obd, struct obd_info *oinfo,
1464                         struct lov_request_set **reqset)
1465 {
1466         struct lov_request_set *set;
1467         struct lov_obd *lov = &obd->u.lov;
1468         int rc = 0, i;
1469         ENTRY;
1470
1471         OBD_ALLOC(set, sizeof(*set));
1472         if (set == NULL)
1473                 RETURN(-ENOMEM);
1474         lov_init_set(set);
1475
1476         set->set_obd = obd;
1477         set->set_oi = oinfo;
1478
1479         /* We only get block data from the OBD */
1480         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1481                 struct lov_request *req;
1482
1483                 if (!lov->lov_tgts[i] || (!lov->lov_tgts[i]->ltd_active
1484                                           && (oinfo->oi_flags & OBD_STATFS_NODELAY))) {
1485                         CDEBUG(D_HA, "lov idx %d inactive\n", i);
1486                         continue;
1487                 }
1488
1489                 /* skip targets that have been explicitely disabled by the
1490                  * administrator */
1491                 if (!lov->lov_tgts[i]->ltd_exp) {
1492                         CDEBUG(D_HA, "lov idx %d administratively disabled\n", i);
1493                         continue;
1494                 }
1495
1496                 OBD_ALLOC(req, sizeof(*req));
1497                 if (req == NULL)
1498                         GOTO(out_set, rc = -ENOMEM);
1499
1500                 OBD_ALLOC(req->rq_oi.oi_osfs, sizeof(*req->rq_oi.oi_osfs));
1501                 if (req->rq_oi.oi_osfs == NULL) {
1502                         OBD_FREE(req, sizeof(*req));
1503                         GOTO(out_set, rc = -ENOMEM);
1504                 }
1505
1506                 req->rq_idx = i;
1507                 req->rq_oi.oi_cb_up = cb_statfs_update;
1508                 req->rq_oi.oi_flags = oinfo->oi_flags;
1509
1510                 lov_set_add_req(req, set);
1511         }
1512         if (!set->set_count)
1513                 GOTO(out_set, rc = -EIO);
1514         *reqset = set;
1515         RETURN(rc);
1516 out_set:
1517         lov_fini_statfs_set(set);
1518         RETURN(rc);
1519 }