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, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LOV
  41
  42 #ifdef __KERNEL__
  43 #include <libcfs/libcfs.h>
  44 #else
  45 #include <liblustre.h>
  46 #endif
  47
  48 #include <obd_class.h>
  49 #include <obd_lov.h>
  50 #include <lustre/lustre_idl.h>
  51
  52 #include "lov_internal.h"
  53
  54 static void lov_init_set(struct lov_request_set *set)
  55 {
  56         set->set_count = 0;
  57         cfs_atomic_set(&set->set_completes, 0);
  58         cfs_atomic_set(&set->set_success, 0);
  59         set->set_cookies = 0;
  60         CFS_INIT_LIST_HEAD(&set->set_list);
  61         cfs_atomic_set(&set->set_refcount, 1);
  62         cfs_waitq_init(&set->set_waitq);
  63         cfs_spin_lock_init(&set->set_lock);
  64 }
  65
  66 void lov_finish_set(struct lov_request_set *set)
  67 {
  68         cfs_list_t *pos, *n;
  69         ENTRY;
  70
  71         LASSERT(set);
  72         cfs_list_for_each_safe(pos, n, &set->set_list) {
  73                 struct lov_request *req = cfs_list_entry(pos,
  74                                                          struct lov_request,
  75                                                          rq_link);
  76                 cfs_list_del_init(&req->rq_link);
  77
  78                 if (req->rq_oi.oi_oa)
  79                         OBDO_FREE(req->rq_oi.oi_oa);
  80                 if (req->rq_oi.oi_md)
  81                         OBD_FREE_LARGE(req->rq_oi.oi_md, req->rq_buflen);
  82                 if (req->rq_oi.oi_osfs)
  83                         OBD_FREE(req->rq_oi.oi_osfs,
  84                                  sizeof(*req->rq_oi.oi_osfs));
  85                 OBD_FREE(req, sizeof(*req));
  86         }
  87
  88         if (set->set_pga) {
  89                 int len = set->set_oabufs * sizeof(*set->set_pga);
  90                 OBD_FREE_LARGE(set->set_pga, len);
  91         }
  92         if (set->set_lockh)
  93                 lov_llh_put(set->set_lockh);
  94
  95         OBD_FREE(set, sizeof(*set));
  96         EXIT;
  97 }
  98
  99 int lov_finished_set(struct lov_request_set *set)
 100 {
 101         int completes = cfs_atomic_read(&set->set_completes);
 102
 103         CDEBUG(D_INFO, "check set %d/%d\n", completes,
 104                set->set_count);
 105         return completes == set->set_count;
 106 }
 107
 108 void lov_update_set(struct lov_request_set *set,
 109                     struct lov_request *req, int rc)
 110 {
 111         req->rq_complete = 1;
 112         req->rq_rc = rc;
 113
 114         cfs_atomic_inc(&set->set_completes);
 115         if (rc == 0)
 116                 cfs_atomic_inc(&set->set_success);
 117
 118         cfs_waitq_signal(&set->set_waitq);
 119 }
 120
 121 int lov_update_common_set(struct lov_request_set *set,
 122                           struct lov_request *req, int rc)
 123 {
 124         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 125         ENTRY;
 126
 127         lov_update_set(set, req, rc);
 128
 129         /* grace error on inactive ost */
 130         if (rc && !(lov->lov_tgts[req->rq_idx] &&
 131                     lov->lov_tgts[req->rq_idx]->ltd_active))
 132                 rc = 0;
 133
 134         /* FIXME in raid1 regime, should return 0 */
 135         RETURN(rc);
 136 }
 137
 138 void lov_set_add_req(struct lov_request *req, struct lov_request_set *set)
 139 {
 140         cfs_list_add_tail(&req->rq_link, &set->set_list);
 141         set->set_count++;
 142         req->rq_rqset = set;
 143 }
 144
 145 extern void osc_update_enqueue(struct lustre_handle *lov_lockhp,
 146                                struct lov_oinfo *loi, int flags,
 147                                struct ost_lvb *lvb, __u32 mode, int rc);
 148
 149 static int lov_update_enqueue_lov(struct obd_export *exp,
 150                                   struct lustre_handle *lov_lockhp,
 151                                   struct lov_oinfo *loi, int flags, int idx,
 152                                   __u64 oid, int rc)
 153 {
 154         struct lov_obd *lov = &exp->exp_obd->u.lov;
 155
 156         if (rc != ELDLM_OK &&
 157             !(rc == ELDLM_LOCK_ABORTED && (flags & LDLM_FL_HAS_INTENT))) {
 158                 memset(lov_lockhp, 0, sizeof(*lov_lockhp));
 159                 if (lov->lov_tgts[idx] && lov->lov_tgts[idx]->ltd_active) {
 160                         /* -EUSERS used by OST to report file contention */
 161                         if (rc != -EINTR && rc != -EUSERS)
 162                                 CERROR("enqueue objid "LPX64" subobj "
 163                                        LPX64" on OST idx %d: rc %d\n",
 164                                        oid, loi->loi_id, loi->loi_ost_idx, rc);
 165                 } else
 166                         rc = ELDLM_OK;
 167         }
 168         return rc;
 169 }
 170
 171 int lov_update_enqueue_set(struct lov_request *req, __u32 mode, int rc)
 172 {
 173         struct lov_request_set *set = req->rq_rqset;
 174         struct lustre_handle *lov_lockhp;
 175         struct obd_info *oi = set->set_oi;
 176         struct lov_oinfo *loi;
 177         ENTRY;
 178
 179         LASSERT(oi != NULL);
 180
 181         lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
 182         loi = oi->oi_md->lsm_oinfo[req->rq_stripe];
 183
 184         /* XXX LOV STACKING: OSC gets a copy, created in lov_prep_enqueue_set
 185          * and that copy can be arbitrarily out of date.
 186          *
 187          * The LOV API is due for a serious rewriting anyways, and this
 188          * can be addressed then. */
 189
 190         lov_stripe_lock(oi->oi_md);
 191         osc_update_enqueue(lov_lockhp, loi, oi->oi_flags,
 192                            &req->rq_oi.oi_md->lsm_oinfo[0]->loi_lvb, mode, rc);
 193         if (rc == ELDLM_LOCK_ABORTED && (oi->oi_flags & LDLM_FL_HAS_INTENT))
 194                 memset(lov_lockhp, 0, sizeof *lov_lockhp);
 195         rc = lov_update_enqueue_lov(set->set_exp, lov_lockhp, loi, oi->oi_flags,
 196                                     req->rq_idx, oi->oi_md->lsm_object_id, rc);
 197         lov_stripe_unlock(oi->oi_md);
 198         lov_update_set(set, req, rc);
 199         RETURN(rc);
 200 }
 201
 202 /* The callback for osc_enqueue that updates lov info for every OSC request. */
 203 static int cb_update_enqueue(void *cookie, int rc)
 204 {
 205         struct obd_info *oinfo = cookie;
 206         struct ldlm_enqueue_info *einfo;
 207         struct lov_request *lovreq;
 208
 209         lovreq = container_of(oinfo, struct lov_request, rq_oi);
 210         einfo = lovreq->rq_rqset->set_ei;
 211         return lov_update_enqueue_set(lovreq, einfo->ei_mode, rc);
 212 }
 213
 214 static int enqueue_done(struct lov_request_set *set, __u32 mode)
 215 {
 216         struct lov_request *req;
 217         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 218         int completes = cfs_atomic_read(&set->set_completes);
 219         int rc = 0;
 220         ENTRY;
 221
 222         /* enqueue/match success, just return */
 223         if (completes && completes == cfs_atomic_read(&set->set_success))
 224                 RETURN(0);
 225
 226         /* cancel enqueued/matched locks */
 227         cfs_list_for_each_entry(req, &set->set_list, rq_link) {
 228                 struct lustre_handle *lov_lockhp;
 229
 230                 if (!req->rq_complete || req->rq_rc)
 231                         continue;
 232
 233                 lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
 234                 LASSERT(lov_lockhp);
 235                 if (!lustre_handle_is_used(lov_lockhp))
 236                         continue;
 237
 238                 rc = obd_cancel(lov->lov_tgts[req->rq_idx]->ltd_exp,
 239                                 req->rq_oi.oi_md, mode, lov_lockhp);
 240                 if (rc && lov->lov_tgts[req->rq_idx] &&
 241                     lov->lov_tgts[req->rq_idx]->ltd_active)
 242                         CERROR("cancelling obdjid "LPX64" on OST "
 243                                "idx %d error: rc = %d\n",
 244                                req->rq_oi.oi_md->lsm_object_id,
 245                                req->rq_idx, rc);
 246         }
 247         if (set->set_lockh)
 248                 lov_llh_put(set->set_lockh);
 249         RETURN(rc);
 250 }
 251
 252 int lov_fini_enqueue_set(struct lov_request_set *set, __u32 mode, int rc,
 253                          struct ptlrpc_request_set *rqset)
 254 {
 255         int ret = 0;
 256         ENTRY;
 257
 258         if (set == NULL)
 259                 RETURN(0);
 260         LASSERT(set->set_exp);
 261         /* Do enqueue_done only for sync requests and if any request
 262          * succeeded. */
 263         if (!rqset) {
 264                 if (rc)
 265                         cfs_atomic_set(&set->set_completes, 0);
 266                 ret = enqueue_done(set, mode);
 267         } else if (set->set_lockh)
 268                 lov_llh_put(set->set_lockh);
 269
 270         lov_put_reqset(set);
 271
 272         RETURN(rc ? rc : ret);
 273 }
 274
 275 static void lov_llh_addref(void *llhp)
 276 {
 277         struct lov_lock_handles *llh = llhp;
 278
 279         cfs_atomic_inc(&llh->llh_refcount);
 280         CDEBUG(D_INFO, "GETting llh %p : new refcount %d\n", llh,
 281                cfs_atomic_read(&llh->llh_refcount));
 282 }
 283
 284 static struct portals_handle_ops lov_handle_ops = {
 285         .hop_addref = lov_llh_addref,
 286         .hop_free   = NULL,
 287 };
 288
 289 static struct lov_lock_handles *lov_llh_new(struct lov_stripe_md *lsm)
 290 {
 291         struct lov_lock_handles *llh;
 292
 293         OBD_ALLOC(llh, sizeof *llh +
 294                   sizeof(*llh->llh_handles) * lsm->lsm_stripe_count);
 295         if (llh == NULL)
 296                 return NULL;
 297
 298         cfs_atomic_set(&llh->llh_refcount, 2);
 299         llh->llh_stripe_count = lsm->lsm_stripe_count;
 300         CFS_INIT_LIST_HEAD(&llh->llh_handle.h_link);
 301         class_handle_hash(&llh->llh_handle, &lov_handle_ops);
 302
 303         return llh;
 304 }
 305
 306 int lov_prep_enqueue_set(struct obd_export *exp, struct obd_info *oinfo,
 307                          struct ldlm_enqueue_info *einfo,
 308                          struct lov_request_set **reqset)
 309 {
 310         struct lov_obd *lov = &exp->exp_obd->u.lov;
 311         struct lov_request_set *set;
 312         int i, rc = 0;
 313         ENTRY;
 314
 315         OBD_ALLOC(set, sizeof(*set));
 316         if (set == NULL)
 317                 RETURN(-ENOMEM);
 318         lov_init_set(set);
 319
 320         set->set_exp = exp;
 321         set->set_oi = oinfo;
 322         set->set_ei = einfo;
 323         set->set_lockh = lov_llh_new(oinfo->oi_md);
 324         if (set->set_lockh == NULL)
 325                 GOTO(out_set, rc = -ENOMEM);
 326         oinfo->oi_lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 327
 328         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
 329                 struct lov_oinfo *loi;
 330                 struct lov_request *req;
 331                 obd_off start, end;
 332
 333                 loi = oinfo->oi_md->lsm_oinfo[i];
 334                 if (!lov_stripe_intersects(oinfo->oi_md, i,
 335                                            oinfo->oi_policy.l_extent.start,
 336                                            oinfo->oi_policy.l_extent.end,
 337                                            &start, &end))
 338                         continue;
 339
 340                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 341                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 342                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 343                         continue;
 344                 }
 345
 346                 OBD_ALLOC(req, sizeof(*req));
 347                 if (req == NULL)
 348                         GOTO(out_set, rc = -ENOMEM);
 349
 350                 req->rq_buflen = sizeof(*req->rq_oi.oi_md) +
 351                         sizeof(struct lov_oinfo *) +
 352                         sizeof(struct lov_oinfo);
 353                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 354                 if (req->rq_oi.oi_md == NULL) {
 355                         OBD_FREE(req, sizeof(*req));
 356                         GOTO(out_set, rc = -ENOMEM);
 357                 }
 358                 req->rq_oi.oi_md->lsm_oinfo[0] =
 359                         ((void *)req->rq_oi.oi_md) + sizeof(*req->rq_oi.oi_md) +
 360                         sizeof(struct lov_oinfo *);
 361
 362                 /* Set lov request specific parameters. */
 363                 req->rq_oi.oi_lockh = set->set_lockh->llh_handles + i;
 364                 req->rq_oi.oi_cb_up = cb_update_enqueue;
 365                 req->rq_oi.oi_flags = oinfo->oi_flags;
 366
 367                 LASSERT(req->rq_oi.oi_lockh);
 368
 369                 req->rq_oi.oi_policy.l_extent.gid =
 370                         oinfo->oi_policy.l_extent.gid;
 371                 req->rq_oi.oi_policy.l_extent.start = start;
 372                 req->rq_oi.oi_policy.l_extent.end = end;
 373
 374                 req->rq_idx = loi->loi_ost_idx;
 375                 req->rq_stripe = i;
 376
 377                 /* XXX LOV STACKING: submd should be from the subobj */
 378                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 379                 req->rq_oi.oi_md->lsm_object_seq = loi->loi_seq;
 380                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 381                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_kms_valid =
 382                         loi->loi_kms_valid;
 383                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_kms = loi->loi_kms;
 384                 req->rq_oi.oi_md->lsm_oinfo[0]->loi_lvb = loi->loi_lvb;
 385
 386                 lov_set_add_req(req, set);
 387         }
 388         if (!set->set_count)
 389                 GOTO(out_set, rc = -EIO);
 390         *reqset = set;
 391         RETURN(0);
 392 out_set:
 393         lov_fini_enqueue_set(set, einfo->ei_mode, rc, NULL);
 394         RETURN(rc);
 395 }
 396
 397 int lov_fini_match_set(struct lov_request_set *set, __u32 mode, int flags)
 398 {
 399         int rc = 0;
 400         ENTRY;
 401
 402         if (set == NULL)
 403                 RETURN(0);
 404         LASSERT(set->set_exp);
 405         rc = enqueue_done(set, mode);
 406         if ((set->set_count == cfs_atomic_read(&set->set_success)) &&
 407             (flags & LDLM_FL_TEST_LOCK))
 408                 lov_llh_put(set->set_lockh);
 409
 410         lov_put_reqset(set);
 411
 412         RETURN(rc);
 413 }
 414
 415 int lov_prep_match_set(struct obd_export *exp, struct obd_info *oinfo,
 416                        struct lov_stripe_md *lsm, ldlm_policy_data_t *policy,
 417                        __u32 mode, struct lustre_handle *lockh,
 418                        struct lov_request_set **reqset)
 419 {
 420         struct lov_obd *lov = &exp->exp_obd->u.lov;
 421         struct lov_request_set *set;
 422         int i, rc = 0;
 423         ENTRY;
 424
 425         OBD_ALLOC(set, sizeof(*set));
 426         if (set == NULL)
 427                 RETURN(-ENOMEM);
 428         lov_init_set(set);
 429
 430         set->set_exp = exp;
 431         set->set_oi = oinfo;
 432         set->set_oi->oi_md = lsm;
 433         set->set_lockh = lov_llh_new(lsm);
 434         if (set->set_lockh == NULL)
 435                 GOTO(out_set, rc = -ENOMEM);
 436         lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 437
 438         for (i = 0; i < lsm->lsm_stripe_count; i++){
 439                 struct lov_oinfo *loi;
 440                 struct lov_request *req;
 441                 obd_off start, end;
 442
 443                 loi = lsm->lsm_oinfo[i];
 444                 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
 445                                            policy->l_extent.end, &start, &end))
 446                         continue;
 447
 448                 /* FIXME raid1 should grace this error */
 449                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 450                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 451                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 452                         GOTO(out_set, rc = -EIO);
 453                 }
 454
 455                 OBD_ALLOC(req, sizeof(*req));
 456                 if (req == NULL)
 457                         GOTO(out_set, rc = -ENOMEM);
 458
 459                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 460                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 461                 if (req->rq_oi.oi_md == NULL) {
 462                         OBD_FREE(req, sizeof(*req));
 463                         GOTO(out_set, rc = -ENOMEM);
 464                 }
 465
 466                 req->rq_oi.oi_policy.l_extent.start = start;
 467                 req->rq_oi.oi_policy.l_extent.end = end;
 468                 req->rq_oi.oi_policy.l_extent.gid = policy->l_extent.gid;
 469
 470                 req->rq_idx = loi->loi_ost_idx;
 471                 req->rq_stripe = i;
 472
 473                 /* XXX LOV STACKING: submd should be from the subobj */
 474                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 475                 req->rq_oi.oi_md->lsm_object_seq = loi->loi_seq;
 476                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 477
 478                 lov_set_add_req(req, set);
 479         }
 480         if (!set->set_count)
 481                 GOTO(out_set, rc = -EIO);
 482         *reqset = set;
 483         RETURN(rc);
 484 out_set:
 485         lov_fini_match_set(set, mode, 0);
 486         RETURN(rc);
 487 }
 488
 489 int lov_fini_cancel_set(struct lov_request_set *set)
 490 {
 491         int rc = 0;
 492         ENTRY;
 493
 494         if (set == NULL)
 495                 RETURN(0);
 496
 497         LASSERT(set->set_exp);
 498         if (set->set_lockh)
 499                 lov_llh_put(set->set_lockh);
 500
 501         lov_put_reqset(set);
 502
 503         RETURN(rc);
 504 }
 505
 506 int lov_prep_cancel_set(struct obd_export *exp, struct obd_info *oinfo,
 507                         struct lov_stripe_md *lsm, __u32 mode,
 508                         struct lustre_handle *lockh,
 509                         struct lov_request_set **reqset)
 510 {
 511         struct lov_request_set *set;
 512         int i, rc = 0;
 513         ENTRY;
 514
 515         OBD_ALLOC(set, sizeof(*set));
 516         if (set == NULL)
 517                 RETURN(-ENOMEM);
 518         lov_init_set(set);
 519
 520         set->set_exp = exp;
 521         set->set_oi = oinfo;
 522         set->set_oi->oi_md = lsm;
 523         set->set_lockh = lov_handle2llh(lockh);
 524         if (set->set_lockh == NULL) {
 525                 CERROR("LOV: invalid lov lock handle %p\n", lockh);
 526                 GOTO(out_set, rc = -EINVAL);
 527         }
 528         lockh->cookie = set->set_lockh->llh_handle.h_cookie;
 529
 530         for (i = 0; i < lsm->lsm_stripe_count; i++){
 531                 struct lov_request *req;
 532                 struct lustre_handle *lov_lockhp;
 533                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 534
 535                 lov_lockhp = set->set_lockh->llh_handles + i;
 536                 if (!lustre_handle_is_used(lov_lockhp)) {
 537                         CDEBUG(D_INFO, "lov idx %d subobj "LPX64" no lock\n",
 538                                loi->loi_ost_idx, loi->loi_id);
 539                         continue;
 540                 }
 541
 542                 OBD_ALLOC(req, sizeof(*req));
 543                 if (req == NULL)
 544                         GOTO(out_set, rc = -ENOMEM);
 545
 546                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 547                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 548                 if (req->rq_oi.oi_md == NULL) {
 549                         OBD_FREE(req, sizeof(*req));
 550                         GOTO(out_set, rc = -ENOMEM);
 551                 }
 552
 553                 req->rq_idx = loi->loi_ost_idx;
 554                 req->rq_stripe = i;
 555
 556                 /* XXX LOV STACKING: submd should be from the subobj */
 557                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 558                 req->rq_oi.oi_md->lsm_object_seq = loi->loi_seq;
 559                 req->rq_oi.oi_md->lsm_stripe_count = 0;
 560
 561                 lov_set_add_req(req, set);
 562         }
 563         if (!set->set_count)
 564                 GOTO(out_set, rc = -EIO);
 565         *reqset = set;
 566         RETURN(rc);
 567 out_set:
 568         lov_fini_cancel_set(set);
 569         RETURN(rc);
 570 }
 571
 572 static int lov_update_create_set(struct lov_request_set *set,
 573                                  struct lov_request *req, int rc)
 574 {
 575         struct obd_trans_info *oti = set->set_oti;
 576         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 577         struct lov_oinfo *loi;
 578         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 579         ENTRY;
 580
 581         if (rc && lov->lov_tgts[req->rq_idx] &&
 582             lov->lov_tgts[req->rq_idx]->ltd_active) {
 583                 /* Pre-creating objects may timeout via -ETIMEDOUT or
 584                  * -ENOTCONN both are always non-critical events. */
 585                 CDEBUG(rc == -ETIMEDOUT || rc == -ENOTCONN ? D_HA : D_ERROR,
 586                        "error creating fid "LPX64" sub-object "
 587                        "on OST idx %d/%d: rc = %d\n",
 588                        set->set_oi->oi_oa->o_id, req->rq_idx,
 589                        lsm->lsm_stripe_count, rc);
 590                 if (rc > 0) {
 591                         CERROR("obd_create returned invalid err %d\n", rc);
 592                         rc = -EIO;
 593                 }
 594         }
 595
 596         cfs_spin_lock(&set->set_lock);
 597         req->rq_stripe = cfs_atomic_read(&set->set_success);
 598         loi = lsm->lsm_oinfo[req->rq_stripe];
 599
 600
 601         if (rc) {
 602                 lov_update_set(set, req, rc);
 603                 cfs_spin_unlock(&set->set_lock);
 604                 RETURN(rc);
 605         }
 606
 607         loi->loi_id = req->rq_oi.oi_oa->o_id;
 608         loi->loi_seq = req->rq_oi.oi_oa->o_seq;
 609         loi->loi_ost_idx = req->rq_idx;
 610         loi_init(loi);
 611
 612         if (oti && set->set_cookies)
 613                 ++oti->oti_logcookies;
 614         if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLCOOKIE)
 615                 set->set_cookie_sent++;
 616
 617         lov_update_set(set, req, rc);
 618         cfs_spin_unlock(&set->set_lock);
 619
 620         CDEBUG(D_INODE, "objid "LPX64" has subobj "LPX64"/"LPU64" at idx %d\n",
 621                lsm->lsm_object_id, loi->loi_id, loi->loi_id, req->rq_idx);
 622         RETURN(rc);
 623 }
 624
 625 static int create_done(struct obd_export *exp, struct lov_request_set *set,
 626                        struct lov_stripe_md **lsmp)
 627 {
 628         struct lov_obd *lov = &exp->exp_obd->u.lov;
 629         struct obd_trans_info *oti = set->set_oti;
 630         struct obdo *src_oa = set->set_oi->oi_oa;
 631         struct lov_request *req;
 632         struct obdo *ret_oa = NULL;
 633         int success, attrset = 0, rc = 0;
 634         ENTRY;
 635
 636         LASSERT(cfs_atomic_read(&set->set_completes));
 637
 638         /* try alloc objects on other osts if osc_create fails for
 639          * exceptions: RPC failure, ENOSPC, etc */
 640         if (set->set_count != cfs_atomic_read(&set->set_success)) {
 641                 cfs_list_for_each_entry (req, &set->set_list, rq_link) {
 642                         if (req->rq_rc == 0)
 643                                 continue;
 644
 645                         cfs_atomic_dec(&set->set_completes);
 646                         req->rq_complete = 0;
 647
 648                         rc = qos_remedy_create(set, req);
 649                         lov_update_create_set(set, req, rc);
 650                 }
 651         }
 652
 653         success = cfs_atomic_read(&set->set_success);
 654         /* no successful creates */
 655         if (success == 0)
 656                 GOTO(cleanup, rc);
 657
 658         if (set->set_count != success) {
 659                 set->set_count = success;
 660                 qos_shrink_lsm(set);
 661         }
 662
 663         OBDO_ALLOC(ret_oa);
 664         if (ret_oa == NULL)
 665                 GOTO(cleanup, rc = -ENOMEM);
 666
 667         cfs_list_for_each_entry(req, &set->set_list, rq_link) {
 668                 if (!req->rq_complete || req->rq_rc)
 669                         continue;
 670                 lov_merge_attrs(ret_oa, req->rq_oi.oi_oa,
 671                                 req->rq_oi.oi_oa->o_valid, set->set_oi->oi_md,
 672                                 req->rq_stripe, &attrset);
 673         }
 674         if (src_oa->o_valid & OBD_MD_FLSIZE &&
 675             ret_oa->o_size != src_oa->o_size) {
 676                 CERROR("original size "LPU64" isn't new object size "LPU64"\n",
 677                        src_oa->o_size, ret_oa->o_size);
 678                 LBUG();
 679         }
 680         ret_oa->o_id = src_oa->o_id;
 681         ret_oa->o_seq = src_oa->o_seq;
 682         ret_oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 683         memcpy(src_oa, ret_oa, sizeof(*src_oa));
 684         OBDO_FREE(ret_oa);
 685
 686         *lsmp = set->set_oi->oi_md;
 687         GOTO(done, rc = 0);
 688
 689 cleanup:
 690         cfs_list_for_each_entry(req, &set->set_list, rq_link) {
 691                 struct obd_export *sub_exp;
 692                 int err = 0;
 693
 694                 if (!req->rq_complete || req->rq_rc)
 695                         continue;
 696
 697                 sub_exp = lov->lov_tgts[req->rq_idx]->ltd_exp;
 698                 err = obd_destroy(NULL, sub_exp, req->rq_oi.oi_oa, NULL, oti,
 699                                   NULL, NULL);
 700                 if (err)
 701                         CERROR("Failed to uncreate objid "LPX64" subobj "
 702                                LPX64" on OST idx %d: rc = %d\n",
 703                                src_oa->o_id, req->rq_oi.oi_oa->o_id,
 704                                req->rq_idx, rc);
 705         }
 706         if (*lsmp == NULL)
 707                 obd_free_memmd(exp, &set->set_oi->oi_md);
 708 done:
 709         if (oti && set->set_cookies) {
 710                 oti->oti_logcookies = set->set_cookies;
 711                 if (!set->set_cookie_sent) {
 712                         oti_free_cookies(oti);
 713                         src_oa->o_valid &= ~OBD_MD_FLCOOKIE;
 714                 } else {
 715                         src_oa->o_valid |= OBD_MD_FLCOOKIE;
 716                 }
 717         }
 718         RETURN(rc);
 719 }
 720
 721 int lov_fini_create_set(struct lov_request_set *set,struct lov_stripe_md **lsmp)
 722 {
 723         int rc = 0;
 724         ENTRY;
 725
 726         if (set == NULL)
 727                 RETURN(0);
 728         LASSERT(set->set_exp);
 729         if (cfs_atomic_read(&set->set_completes))
 730                 rc = create_done(set->set_exp, set, lsmp);
 731
 732         lov_put_reqset(set);
 733         RETURN(rc);
 734 }
 735
 736 int cb_create_update(void *cookie, int rc)
 737 {
 738         struct obd_info *oinfo = cookie;
 739         struct lov_request *lovreq;
 740
 741         lovreq = container_of(oinfo, struct lov_request, rq_oi);
 742
 743         if (CFS_FAIL_CHECK(OBD_FAIL_MDS_OSC_CREATE_FAIL))
 744                 if (lovreq->rq_idx == cfs_fail_val)
 745                         rc = -ENOTCONN;
 746
 747         rc= lov_update_create_set(lovreq->rq_rqset, lovreq, rc);
 748         if (lov_finished_set(lovreq->rq_rqset))
 749                 lov_put_reqset(lovreq->rq_rqset);
 750         return rc;
 751 }
 752
 753 int lov_prep_create_set(struct obd_export *exp, struct obd_info *oinfo,
 754                         struct lov_stripe_md **lsmp, struct obdo *src_oa,
 755                         struct obd_trans_info *oti,
 756                         struct lov_request_set **reqset)
 757 {
 758         struct lov_request_set *set;
 759         int rc = 0;
 760         ENTRY;
 761
 762         OBD_ALLOC(set, sizeof(*set));
 763         if (set == NULL)
 764                 RETURN(-ENOMEM);
 765         lov_init_set(set);
 766
 767         set->set_exp = exp;
 768         set->set_oi = oinfo;
 769         set->set_oi->oi_md = *lsmp;
 770         set->set_oi->oi_oa = src_oa;
 771         set->set_oti = oti;
 772         lov_get_reqset(set);
 773
 774         rc = qos_prep_create(exp, set);
 775         /* qos_shrink_lsm() may have allocated a new lsm */
 776         *lsmp = oinfo->oi_md;
 777         if (rc) {
 778                 lov_fini_create_set(set, lsmp);
 779                 lov_put_reqset(set);
 780         } else {
 781                 *reqset = set;
 782         }
 783         RETURN(rc);
 784 }
 785
 786 static int common_attr_done(struct lov_request_set *set)
 787 {
 788         cfs_list_t *pos;
 789         struct lov_request *req;
 790         struct obdo *tmp_oa;
 791         int rc = 0, attrset = 0;
 792         ENTRY;
 793
 794         LASSERT(set->set_oi != NULL);
 795
 796         if (set->set_oi->oi_oa == NULL)
 797                 RETURN(0);
 798
 799         if (!cfs_atomic_read(&set->set_success))
 800                 RETURN(-EIO);
 801
 802         OBDO_ALLOC(tmp_oa);
 803         if (tmp_oa == NULL)
 804                 GOTO(out, rc = -ENOMEM);
 805
 806         cfs_list_for_each (pos, &set->set_list) {
 807                 req = cfs_list_entry(pos, struct lov_request, rq_link);
 808
 809                 if (!req->rq_complete || req->rq_rc)
 810                         continue;
 811                 if (req->rq_oi.oi_oa->o_valid == 0)   /* inactive stripe */
 812                         continue;
 813                 lov_merge_attrs(tmp_oa, req->rq_oi.oi_oa,
 814                                 req->rq_oi.oi_oa->o_valid,
 815                                 set->set_oi->oi_md, req->rq_stripe, &attrset);
 816         }
 817         if (!attrset) {
 818                 CERROR("No stripes had valid attrs\n");
 819                 rc = -EIO;
 820         }
 821         if ((set->set_oi->oi_oa->o_valid & OBD_MD_FLEPOCH) &&
 822             (set->set_oi->oi_md->lsm_stripe_count != attrset)) {
 823                 /* When we take attributes of some epoch, we require all the
 824                  * ost to be active. */
 825                 CERROR("Not all the stripes had valid attrs\n");
 826                 GOTO(out, rc = -EIO);
 827         }
 828
 829         tmp_oa->o_id = set->set_oi->oi_oa->o_id;
 830         memcpy(set->set_oi->oi_oa, tmp_oa, sizeof(*set->set_oi->oi_oa));
 831 out:
 832         if (tmp_oa)
 833                 OBDO_FREE(tmp_oa);
 834         RETURN(rc);
 835
 836 }
 837
 838 static int brw_done(struct lov_request_set *set)
 839 {
 840         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 841         struct lov_oinfo     *loi = NULL;
 842         cfs_list_t *pos;
 843         struct lov_request *req;
 844         ENTRY;
 845
 846         cfs_list_for_each (pos, &set->set_list) {
 847                 req = cfs_list_entry(pos, struct lov_request, rq_link);
 848
 849                 if (!req->rq_complete || req->rq_rc)
 850                         continue;
 851
 852                 loi = lsm->lsm_oinfo[req->rq_stripe];
 853
 854                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLBLOCKS)
 855                         loi->loi_lvb.lvb_blocks = req->rq_oi.oi_oa->o_blocks;
 856         }
 857
 858         RETURN(0);
 859 }
 860
 861 int lov_fini_brw_set(struct lov_request_set *set)
 862 {
 863         int rc = 0;
 864         ENTRY;
 865
 866         if (set == NULL)
 867                 RETURN(0);
 868         LASSERT(set->set_exp);
 869         if (cfs_atomic_read(&set->set_completes)) {
 870                 rc = brw_done(set);
 871                 /* FIXME update qos data here */
 872         }
 873         lov_put_reqset(set);
 874
 875         RETURN(rc);
 876 }
 877
 878 int lov_prep_brw_set(struct obd_export *exp, struct obd_info *oinfo,
 879                      obd_count oa_bufs, struct brw_page *pga,
 880                      struct obd_trans_info *oti,
 881                      struct lov_request_set **reqset)
 882 {
 883         struct {
 884                 obd_count       index;
 885                 obd_count       count;
 886                 obd_count       off;
 887         } *info = NULL;
 888         struct lov_request_set *set;
 889         struct lov_obd *lov = &exp->exp_obd->u.lov;
 890         int rc = 0, i, shift;
 891         ENTRY;
 892
 893         OBD_ALLOC(set, sizeof(*set));
 894         if (set == NULL)
 895                 RETURN(-ENOMEM);
 896         lov_init_set(set);
 897
 898         set->set_exp = exp;
 899         set->set_oti = oti;
 900         set->set_oi = oinfo;
 901         set->set_oabufs = oa_bufs;
 902         OBD_ALLOC_LARGE(set->set_pga, oa_bufs * sizeof(*set->set_pga));
 903         if (!set->set_pga)
 904                 GOTO(out, rc = -ENOMEM);
 905
 906         OBD_ALLOC_LARGE(info, sizeof(*info) * oinfo->oi_md->lsm_stripe_count);
 907         if (!info)
 908                 GOTO(out, rc = -ENOMEM);
 909
 910         /* calculate the page count for each stripe */
 911         for (i = 0; i < oa_bufs; i++) {
 912                 int stripe = lov_stripe_number(oinfo->oi_md, pga[i].off);
 913                 info[stripe].count++;
 914         }
 915
 916         /* alloc and initialize lov request */
 917         shift = 0;
 918         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++){
 919                 struct lov_oinfo *loi = NULL;
 920                 struct lov_request *req;
 921
 922                 if (info[i].count == 0)
 923                         continue;
 924
 925                 loi = oinfo->oi_md->lsm_oinfo[i];
 926                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
 927                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
 928                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 929                         GOTO(out, rc = -EIO);
 930                 }
 931
 932                 OBD_ALLOC(req, sizeof(*req));
 933                 if (req == NULL)
 934                         GOTO(out, rc = -ENOMEM);
 935
 936                 OBDO_ALLOC(req->rq_oi.oi_oa);
 937                 if (req->rq_oi.oi_oa == NULL) {
 938                         OBD_FREE(req, sizeof(*req));
 939                         GOTO(out, rc = -ENOMEM);
 940                 }
 941
 942                 if (oinfo->oi_oa) {
 943                         memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
 944                                sizeof(*req->rq_oi.oi_oa));
 945                 }
 946                 req->rq_oi.oi_oa->o_id = loi->loi_id;
 947                 req->rq_oi.oi_oa->o_seq = loi->loi_seq;
 948                 req->rq_oi.oi_oa->o_stripe_idx = i;
 949
 950                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 951                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
 952                 if (req->rq_oi.oi_md == NULL) {
 953                         OBDO_FREE(req->rq_oi.oi_oa);
 954                         OBD_FREE(req, sizeof(*req));
 955                         GOTO(out, rc = -ENOMEM);
 956                 }
 957
 958                 req->rq_idx = loi->loi_ost_idx;
 959                 req->rq_stripe = i;
 960
 961                 /* XXX LOV STACKING */
 962                 req->rq_oi.oi_md->lsm_object_id = loi->loi_id;
 963                 req->rq_oi.oi_md->lsm_object_seq = loi->loi_seq;
 964                 req->rq_oabufs = info[i].count;
 965                 req->rq_pgaidx = shift;
 966                 shift += req->rq_oabufs;
 967
 968                 /* remember the index for sort brw_page array */
 969                 info[i].index = req->rq_pgaidx;
 970
 971                 req->rq_oi.oi_capa = oinfo->oi_capa;
 972
 973                 lov_set_add_req(req, set);
 974         }
 975         if (!set->set_count)
 976                 GOTO(out, rc = -EIO);
 977
 978         /* rotate & sort the brw_page array */
 979         for (i = 0; i < oa_bufs; i++) {
 980                 int stripe = lov_stripe_number(oinfo->oi_md, pga[i].off);
 981
 982                 shift = info[stripe].index + info[stripe].off;
 983                 LASSERT(shift < oa_bufs);
 984                 set->set_pga[shift] = pga[i];
 985                 lov_stripe_offset(oinfo->oi_md, pga[i].off, stripe,
 986                                   &set->set_pga[shift].off);
 987                 info[stripe].off++;
 988         }
 989 out:
 990         if (info)
 991                 OBD_FREE_LARGE(info,
 992                                sizeof(*info) * oinfo->oi_md->lsm_stripe_count);
 993
 994         if (rc == 0)
 995                 *reqset = set;
 996         else
 997                 lov_fini_brw_set(set);
 998
 999         RETURN(rc);
1000 }
1001
1002 int lov_fini_getattr_set(struct lov_request_set *set)
1003 {
1004         int rc = 0;
1005         ENTRY;
1006
1007         if (set == NULL)
1008                 RETURN(0);
1009         LASSERT(set->set_exp);
1010         if (cfs_atomic_read(&set->set_completes))
1011                 rc = common_attr_done(set);
1012
1013         lov_put_reqset(set);
1014
1015         RETURN(rc);
1016 }
1017
1018 /* The callback for osc_getattr_async that finilizes a request info when a
1019  * response is received. */
1020 static int cb_getattr_update(void *cookie, int rc)
1021 {
1022         struct obd_info *oinfo = cookie;
1023         struct lov_request *lovreq;
1024         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1025         return lov_update_common_set(lovreq->rq_rqset, lovreq, rc);
1026 }
1027
1028 int lov_prep_getattr_set(struct obd_export *exp, struct obd_info *oinfo,
1029                          struct lov_request_set **reqset)
1030 {
1031         struct lov_request_set *set;
1032         struct lov_obd *lov = &exp->exp_obd->u.lov;
1033         int rc = 0, i;
1034         ENTRY;
1035
1036         OBD_ALLOC(set, sizeof(*set));
1037         if (set == NULL)
1038                 RETURN(-ENOMEM);
1039         lov_init_set(set);
1040
1041         set->set_exp = exp;
1042         set->set_oi = oinfo;
1043
1044         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1045                 struct lov_oinfo *loi;
1046                 struct lov_request *req;
1047
1048                 loi = oinfo->oi_md->lsm_oinfo[i];
1049                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1050                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1051                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1052                         if (oinfo->oi_oa->o_valid & OBD_MD_FLEPOCH)
1053                                 /* SOM requires all the OSTs to be active. */
1054                                 GOTO(out_set, rc = -EIO);
1055                         continue;
1056                 }
1057
1058                 OBD_ALLOC(req, sizeof(*req));
1059                 if (req == NULL)
1060                         GOTO(out_set, rc = -ENOMEM);
1061
1062                 req->rq_stripe = i;
1063                 req->rq_idx = loi->loi_ost_idx;
1064
1065                 OBDO_ALLOC(req->rq_oi.oi_oa);
1066                 if (req->rq_oi.oi_oa == NULL) {
1067                         OBD_FREE(req, sizeof(*req));
1068                         GOTO(out_set, rc = -ENOMEM);
1069                 }
1070                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1071                        sizeof(*req->rq_oi.oi_oa));
1072                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1073                 req->rq_oi.oi_oa->o_seq = loi->loi_seq;
1074                 req->rq_oi.oi_cb_up = cb_getattr_update;
1075                 req->rq_oi.oi_capa = oinfo->oi_capa;
1076
1077                 lov_set_add_req(req, set);
1078         }
1079         if (!set->set_count)
1080                 GOTO(out_set, rc = -EIO);
1081         *reqset = set;
1082         RETURN(rc);
1083 out_set:
1084         lov_fini_getattr_set(set);
1085         RETURN(rc);
1086 }
1087
1088 int lov_fini_destroy_set(struct lov_request_set *set)
1089 {
1090         ENTRY;
1091
1092         if (set == NULL)
1093                 RETURN(0);
1094         LASSERT(set->set_exp);
1095         if (cfs_atomic_read(&set->set_completes)) {
1096                 /* FIXME update qos data here */
1097         }
1098
1099         lov_put_reqset(set);
1100
1101         RETURN(0);
1102 }
1103
1104 int lov_prep_destroy_set(struct obd_export *exp, struct obd_info *oinfo,
1105                          struct obdo *src_oa, struct lov_stripe_md *lsm,
1106                          struct obd_trans_info *oti,
1107                          struct lov_request_set **reqset)
1108 {
1109         struct lov_request_set *set;
1110         struct lov_obd *lov = &exp->exp_obd->u.lov;
1111         int rc = 0, i;
1112         ENTRY;
1113
1114         OBD_ALLOC(set, sizeof(*set));
1115         if (set == NULL)
1116                 RETURN(-ENOMEM);
1117         lov_init_set(set);
1118
1119         set->set_exp = exp;
1120         set->set_oi = oinfo;
1121         set->set_oi->oi_md = lsm;
1122         set->set_oi->oi_oa = src_oa;
1123         set->set_oti = oti;
1124         if (oti != NULL && src_oa->o_valid & OBD_MD_FLCOOKIE)
1125                 set->set_cookies = oti->oti_logcookies;
1126
1127         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1128                 struct lov_oinfo *loi;
1129                 struct lov_request *req;
1130
1131                 loi = lsm->lsm_oinfo[i];
1132                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1133                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1134                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1135                         continue;
1136                 }
1137
1138                 OBD_ALLOC(req, sizeof(*req));
1139                 if (req == NULL)
1140                         GOTO(out_set, rc = -ENOMEM);
1141
1142                 req->rq_stripe = i;
1143                 req->rq_idx = loi->loi_ost_idx;
1144
1145                 OBDO_ALLOC(req->rq_oi.oi_oa);
1146                 if (req->rq_oi.oi_oa == NULL) {
1147                         OBD_FREE(req, sizeof(*req));
1148                         GOTO(out_set, rc = -ENOMEM);
1149                 }
1150                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1151                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1152                 req->rq_oi.oi_oa->o_seq = loi->loi_seq;
1153                 lov_set_add_req(req, set);
1154         }
1155         if (!set->set_count)
1156                 GOTO(out_set, rc = -EIO);
1157         *reqset = set;
1158         RETURN(rc);
1159 out_set:
1160         lov_fini_destroy_set(set);
1161         RETURN(rc);
1162 }
1163
1164 int lov_fini_setattr_set(struct lov_request_set *set)
1165 {
1166         int rc = 0;
1167         ENTRY;
1168
1169         if (set == NULL)
1170                 RETURN(0);
1171         LASSERT(set->set_exp);
1172         if (cfs_atomic_read(&set->set_completes)) {
1173                 rc = common_attr_done(set);
1174                 /* FIXME update qos data here */
1175         }
1176
1177         lov_put_reqset(set);
1178         RETURN(rc);
1179 }
1180
1181 int lov_update_setattr_set(struct lov_request_set *set,
1182                            struct lov_request *req, int rc)
1183 {
1184         struct lov_obd *lov = &req->rq_rqset->set_exp->exp_obd->u.lov;
1185         struct lov_stripe_md *lsm = req->rq_rqset->set_oi->oi_md;
1186         ENTRY;
1187
1188         lov_update_set(set, req, rc);
1189
1190         /* grace error on inactive ost */
1191         if (rc && !(lov->lov_tgts[req->rq_idx] &&
1192                     lov->lov_tgts[req->rq_idx]->ltd_active))
1193                 rc = 0;
1194
1195         if (rc == 0) {
1196                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLCTIME)
1197                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_ctime =
1198                                 req->rq_oi.oi_oa->o_ctime;
1199                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLMTIME)
1200                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_mtime =
1201                                 req->rq_oi.oi_oa->o_mtime;
1202                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLATIME)
1203                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_atime =
1204                                 req->rq_oi.oi_oa->o_atime;
1205         }
1206
1207         RETURN(rc);
1208 }
1209
1210 /* The callback for osc_setattr_async that finilizes a request info when a
1211  * response is received. */
1212 static int cb_setattr_update(void *cookie, int rc)
1213 {
1214         struct obd_info *oinfo = cookie;
1215         struct lov_request *lovreq;
1216         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1217         return lov_update_setattr_set(lovreq->rq_rqset, lovreq, rc);
1218 }
1219
1220 int lov_prep_setattr_set(struct obd_export *exp, struct obd_info *oinfo,
1221                          struct obd_trans_info *oti,
1222                          struct lov_request_set **reqset)
1223 {
1224         struct lov_request_set *set;
1225         struct lov_obd *lov = &exp->exp_obd->u.lov;
1226         int rc = 0, i;
1227         ENTRY;
1228
1229         OBD_ALLOC(set, sizeof(*set));
1230         if (set == NULL)
1231                 RETURN(-ENOMEM);
1232         lov_init_set(set);
1233
1234         set->set_exp = exp;
1235         set->set_oti = oti;
1236         set->set_oi = oinfo;
1237         if (oti != NULL && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
1238                 set->set_cookies = oti->oti_logcookies;
1239
1240         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1241                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1242                 struct lov_request *req;
1243
1244                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1245                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1246                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1247                         continue;
1248                 }
1249
1250                 OBD_ALLOC(req, sizeof(*req));
1251                 if (req == NULL)
1252                         GOTO(out_set, rc = -ENOMEM);
1253                 req->rq_stripe = i;
1254                 req->rq_idx = loi->loi_ost_idx;
1255
1256                 OBDO_ALLOC(req->rq_oi.oi_oa);
1257                 if (req->rq_oi.oi_oa == NULL) {
1258                         OBD_FREE(req, sizeof(*req));
1259                         GOTO(out_set, rc = -ENOMEM);
1260                 }
1261                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1262                        sizeof(*req->rq_oi.oi_oa));
1263                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1264                 req->rq_oi.oi_oa->o_seq= loi->loi_seq;
1265                 req->rq_oi.oi_oa->o_stripe_idx = i;
1266                 req->rq_oi.oi_cb_up = cb_setattr_update;
1267                 req->rq_oi.oi_capa = oinfo->oi_capa;
1268
1269                 if (oinfo->oi_oa->o_valid & OBD_MD_FLSIZE) {
1270                         int off = lov_stripe_offset(oinfo->oi_md,
1271                                                     oinfo->oi_oa->o_size, i,
1272                                                     &req->rq_oi.oi_oa->o_size);
1273
1274                         if (off < 0 && req->rq_oi.oi_oa->o_size)
1275                                 req->rq_oi.oi_oa->o_size--;
1276
1277                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1278                                i, req->rq_oi.oi_oa->o_size,
1279                                oinfo->oi_oa->o_size);
1280                 }
1281                 lov_set_add_req(req, set);
1282         }
1283         if (!set->set_count)
1284                 GOTO(out_set, rc = -EIO);
1285         *reqset = set;
1286         RETURN(rc);
1287 out_set:
1288         lov_fini_setattr_set(set);
1289         RETURN(rc);
1290 }
1291
1292 int lov_fini_punch_set(struct lov_request_set *set)
1293 {
1294         int rc = 0;
1295         ENTRY;
1296
1297         if (set == NULL)
1298                 RETURN(0);
1299         LASSERT(set->set_exp);
1300         if (cfs_atomic_read(&set->set_completes)) {
1301                 rc = -EIO;
1302                 /* FIXME update qos data here */
1303                 if (cfs_atomic_read(&set->set_success))
1304                         rc = common_attr_done(set);
1305         }
1306
1307         lov_put_reqset(set);
1308
1309         RETURN(rc);
1310 }
1311
1312 int lov_update_punch_set(struct lov_request_set *set,
1313                          struct lov_request *req, int rc)
1314 {
1315         struct lov_obd *lov = &req->rq_rqset->set_exp->exp_obd->u.lov;
1316         struct lov_stripe_md *lsm = req->rq_rqset->set_oi->oi_md;
1317         ENTRY;
1318
1319         lov_update_set(set, req, rc);
1320
1321         /* grace error on inactive ost */
1322         if (rc && !lov->lov_tgts[req->rq_idx]->ltd_active)
1323                 rc = 0;
1324
1325         if (rc == 0) {
1326                 lov_stripe_lock(lsm);
1327                 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLBLOCKS) {
1328                         lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_blocks =
1329                                 req->rq_oi.oi_oa->o_blocks;
1330                 }
1331
1332                 /* Do we need to update lvb_size here? It needn't because
1333                  * it have been done in ll_truncate(). -jay */
1334                 lov_stripe_unlock(lsm);
1335         }
1336
1337         RETURN(rc);
1338 }
1339
1340 /* The callback for osc_punch that finilizes a request info when a response
1341  * is received. */
1342 static int cb_update_punch(void *cookie, int rc)
1343 {
1344         struct obd_info *oinfo = cookie;
1345         struct lov_request *lovreq;
1346         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1347         return lov_update_punch_set(lovreq->rq_rqset, lovreq, rc);
1348 }
1349
1350 int lov_prep_punch_set(struct obd_export *exp, struct obd_info *oinfo,
1351                        struct obd_trans_info *oti,
1352                        struct lov_request_set **reqset)
1353 {
1354         struct lov_request_set *set;
1355         struct lov_obd *lov = &exp->exp_obd->u.lov;
1356         int rc = 0, i;
1357         ENTRY;
1358
1359         OBD_ALLOC(set, sizeof(*set));
1360         if (set == NULL)
1361                 RETURN(-ENOMEM);
1362         lov_init_set(set);
1363
1364         set->set_oi = oinfo;
1365         set->set_exp = exp;
1366
1367         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1368                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1369                 struct lov_request *req;
1370                 obd_off rs, re;
1371
1372                 if (!lov_stripe_intersects(oinfo->oi_md, i,
1373                                            oinfo->oi_policy.l_extent.start,
1374                                            oinfo->oi_policy.l_extent.end,
1375                                            &rs, &re))
1376                         continue;
1377
1378                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1379                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1380                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1381                         GOTO(out_set, rc = -EIO);
1382                 }
1383
1384                 OBD_ALLOC(req, sizeof(*req));
1385                 if (req == NULL)
1386                         GOTO(out_set, rc = -ENOMEM);
1387                 req->rq_stripe = i;
1388                 req->rq_idx = loi->loi_ost_idx;
1389
1390                 OBDO_ALLOC(req->rq_oi.oi_oa);
1391                 if (req->rq_oi.oi_oa == NULL) {
1392                         OBD_FREE(req, sizeof(*req));
1393                         GOTO(out_set, rc = -ENOMEM);
1394                 }
1395                 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
1396                        sizeof(*req->rq_oi.oi_oa));
1397                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1398                 req->rq_oi.oi_oa->o_seq = loi->loi_seq;
1399                 req->rq_oi.oi_oa->o_valid |= OBD_MD_FLGROUP;
1400
1401                 req->rq_oi.oi_oa->o_stripe_idx = i;
1402                 req->rq_oi.oi_cb_up = cb_update_punch;
1403
1404                 req->rq_oi.oi_policy.l_extent.start = rs;
1405                 req->rq_oi.oi_policy.l_extent.end = re;
1406                 req->rq_oi.oi_policy.l_extent.gid = -1;
1407
1408                 req->rq_oi.oi_capa = oinfo->oi_capa;
1409
1410                 lov_set_add_req(req, set);
1411         }
1412         if (!set->set_count)
1413                 GOTO(out_set, rc = -EIO);
1414         *reqset = set;
1415         RETURN(rc);
1416 out_set:
1417         lov_fini_punch_set(set);
1418         RETURN(rc);
1419 }
1420
1421 int lov_fini_sync_set(struct lov_request_set *set)
1422 {
1423         int rc = 0;
1424         ENTRY;
1425
1426         if (set == NULL)
1427                 RETURN(0);
1428         LASSERT(set->set_exp);
1429         if (cfs_atomic_read(&set->set_completes)) {
1430                 if (!cfs_atomic_read(&set->set_success))
1431                         rc = -EIO;
1432                 /* FIXME update qos data here */
1433         }
1434
1435         lov_put_reqset(set);
1436
1437         RETURN(rc);
1438 }
1439
1440 /* The callback for osc_sync that finilizes a request info when a
1441  * response is recieved. */
1442 static int cb_sync_update(void *cookie, int rc)
1443 {
1444         struct obd_info *oinfo = cookie;
1445         struct lov_request *lovreq;
1446
1447         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1448         return lov_update_common_set(lovreq->rq_rqset, lovreq, rc);
1449 }
1450
1451 int lov_prep_sync_set(struct obd_export *exp, struct obd_info *oinfo,
1452                       obd_off start, obd_off end,
1453                       struct lov_request_set **reqset)
1454 {
1455         struct lov_request_set *set;
1456         struct lov_obd *lov = &exp->exp_obd->u.lov;
1457         int rc = 0, i;
1458         ENTRY;
1459
1460         OBD_ALLOC_PTR(set);
1461         if (set == NULL)
1462                 RETURN(-ENOMEM);
1463         lov_init_set(set);
1464
1465         set->set_exp = exp;
1466         set->set_oi = oinfo;
1467
1468         for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
1469                 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
1470                 struct lov_request *req;
1471                 obd_off rs, re;
1472
1473                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1474                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1475                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1476                         continue;
1477                 }
1478
1479                 if (!lov_stripe_intersects(oinfo->oi_md, i, start, end, &rs,
1480                                            &re))
1481                         continue;
1482
1483                 OBD_ALLOC_PTR(req);
1484                 if (req == NULL)
1485                         GOTO(out_set, rc = -ENOMEM);
1486                 req->rq_stripe = i;
1487                 req->rq_idx = loi->loi_ost_idx;
1488
1489                 OBDO_ALLOC(req->rq_oi.oi_oa);
1490                 if (req->rq_oi.oi_oa == NULL) {
1491                         OBD_FREE(req, sizeof(*req));
1492                         GOTO(out_set, rc = -ENOMEM);
1493                 }
1494                 *req->rq_oi.oi_oa = *oinfo->oi_oa;
1495                 req->rq_oi.oi_oa->o_id = loi->loi_id;
1496                 req->rq_oi.oi_oa->o_seq = loi->loi_seq;
1497                 req->rq_oi.oi_oa->o_stripe_idx = i;
1498
1499                 req->rq_oi.oi_policy.l_extent.start = rs;
1500                 req->rq_oi.oi_policy.l_extent.end = re;
1501                 req->rq_oi.oi_policy.l_extent.gid = -1;
1502                 req->rq_oi.oi_cb_up = cb_sync_update;
1503
1504                 lov_set_add_req(req, set);
1505         }
1506         if (!set->set_count)
1507                 GOTO(out_set, rc = -EIO);
1508         *reqset = set;
1509         RETURN(rc);
1510 out_set:
1511         lov_fini_sync_set(set);
1512         RETURN(rc);
1513 }
1514
1515 #define LOV_U64_MAX ((__u64)~0ULL)
1516 #define LOV_SUM_MAX(tot, add)                                           \
1517         do {                                                            \
1518                 if ((tot) + (add) < (tot))                              \
1519                         (tot) = LOV_U64_MAX;                            \
1520                 else                                                    \
1521                         (tot) += (add);                                 \
1522         } while(0)
1523
1524 int lov_fini_statfs(struct obd_device *obd, struct obd_statfs *osfs,int success)
1525 {
1526         ENTRY;
1527
1528         if (success) {
1529                 __u32 expected_stripes = lov_get_stripecnt(&obd->u.lov,
1530                                                            LOV_MAGIC, 0);
1531                 if (osfs->os_files != LOV_U64_MAX)
1532                         lov_do_div64(osfs->os_files, expected_stripes);
1533                 if (osfs->os_ffree != LOV_U64_MAX)
1534                         lov_do_div64(osfs->os_ffree, expected_stripes);
1535
1536                 cfs_spin_lock(&obd->obd_osfs_lock);
1537                 memcpy(&obd->obd_osfs, osfs, sizeof(*osfs));
1538                 obd->obd_osfs_age = cfs_time_current_64();
1539                 cfs_spin_unlock(&obd->obd_osfs_lock);
1540                 RETURN(0);
1541         }
1542
1543         RETURN(-EIO);
1544 }
1545
1546 int lov_fini_statfs_set(struct lov_request_set *set)
1547 {
1548         int rc = 0;
1549         ENTRY;
1550
1551         if (set == NULL)
1552                 RETURN(0);
1553
1554         if (cfs_atomic_read(&set->set_completes)) {
1555                 rc = lov_fini_statfs(set->set_obd, set->set_oi->oi_osfs,
1556                                      cfs_atomic_read(&set->set_success));
1557         }
1558         lov_put_reqset(set);
1559         RETURN(rc);
1560 }
1561
1562 void lov_update_statfs(struct obd_statfs *osfs, struct obd_statfs *lov_sfs,
1563                        int success)
1564 {
1565         int shift = 0, quit = 0;
1566         __u64 tmp;
1567
1568         if (success == 0) {
1569                 memcpy(osfs, lov_sfs, sizeof(*lov_sfs));
1570         } else {
1571                 if (osfs->os_bsize != lov_sfs->os_bsize) {
1572                         /* assume all block sizes are always powers of 2 */
1573                         /* get the bits difference */
1574                         tmp = osfs->os_bsize | lov_sfs->os_bsize;
1575                         for (shift = 0; shift <= 64; ++shift) {
1576                                 if (tmp & 1) {
1577                                         if (quit)
1578                                                 break;
1579                                         else
1580                                                 quit = 1;
1581                                         shift = 0;
1582                                 }
1583                                 tmp >>= 1;
1584                         }
1585                 }
1586
1587                 if (osfs->os_bsize < lov_sfs->os_bsize) {
1588                         osfs->os_bsize = lov_sfs->os_bsize;
1589
1590                         osfs->os_bfree  >>= shift;
1591                         osfs->os_bavail >>= shift;
1592                         osfs->os_blocks >>= shift;
1593                 } else if (shift != 0) {
1594                         lov_sfs->os_bfree  >>= shift;
1595                         lov_sfs->os_bavail >>= shift;
1596                         lov_sfs->os_blocks >>= shift;
1597                 }
1598 #ifdef MIN_DF
1599                 /* Sandia requested that df (and so, statfs) only
1600                    returned minimal available space on
1601                    a single OST, so people would be able to
1602                    write this much data guaranteed. */
1603                 if (osfs->os_bavail > lov_sfs->os_bavail) {
1604                         /* Presumably if new bavail is smaller,
1605                            new bfree is bigger as well */
1606                         osfs->os_bfree = lov_sfs->os_bfree;
1607                         osfs->os_bavail = lov_sfs->os_bavail;
1608                 }
1609 #else
1610                 osfs->os_bfree += lov_sfs->os_bfree;
1611                 osfs->os_bavail += lov_sfs->os_bavail;
1612 #endif
1613                 osfs->os_blocks += lov_sfs->os_blocks;
1614                 /* XXX not sure about this one - depends on policy.
1615                  *   - could be minimum if we always stripe on all OBDs
1616                  *     (but that would be wrong for any other policy,
1617                  *     if one of the OBDs has no more objects left)
1618                  *   - could be sum if we stripe whole objects
1619                  *   - could be average, just to give a nice number
1620                  *
1621                  * To give a "reasonable" (if not wholly accurate)
1622                  * number, we divide the total number of free objects
1623                  * by expected stripe count (watch out for overflow).
1624                  */
1625                 LOV_SUM_MAX(osfs->os_files, lov_sfs->os_files);
1626                 LOV_SUM_MAX(osfs->os_ffree, lov_sfs->os_ffree);
1627         }
1628 }
1629
1630 /* The callback for osc_statfs_async that finilizes a request info when a
1631  * response is received. */
1632 static int cb_statfs_update(void *cookie, int rc)
1633 {
1634         struct obd_info *oinfo = cookie;
1635         struct lov_request *lovreq;
1636         struct lov_request_set *set;
1637         struct obd_statfs *osfs, *lov_sfs;
1638         struct lov_obd *lov;
1639         struct lov_tgt_desc *tgt;
1640         struct obd_device *lovobd, *tgtobd;
1641         int success;
1642         ENTRY;
1643
1644         lovreq = container_of(oinfo, struct lov_request, rq_oi);
1645         set = lovreq->rq_rqset;
1646         lovobd = set->set_obd;
1647         lov = &lovobd->u.lov;
1648         osfs = set->set_oi->oi_osfs;
1649         lov_sfs = oinfo->oi_osfs;
1650         success = cfs_atomic_read(&set->set_success);
1651         /* XXX: the same is done in lov_update_common_set, however
1652            lovset->set_exp is not initialized. */
1653         lov_update_set(set, lovreq, rc);
1654         if (rc)
1655                 GOTO(out, rc);
1656
1657         obd_getref(lovobd);
1658         tgt = lov->lov_tgts[lovreq->rq_idx];
1659         if (!tgt || !tgt->ltd_active)
1660                 GOTO(out_update, rc);
1661
1662         tgtobd = class_exp2obd(tgt->ltd_exp);
1663         cfs_spin_lock(&tgtobd->obd_osfs_lock);
1664         memcpy(&tgtobd->obd_osfs, lov_sfs, sizeof(*lov_sfs));
1665         if ((oinfo->oi_flags & OBD_STATFS_FROM_CACHE) == 0)
1666                 tgtobd->obd_osfs_age = cfs_time_current_64();
1667         cfs_spin_unlock(&tgtobd->obd_osfs_lock);
1668
1669 out_update:
1670         lov_update_statfs(osfs, lov_sfs, success);
1671         qos_update(lov);
1672         obd_putref(lovobd);
1673
1674 out:
1675         if (set->set_oi->oi_flags & OBD_STATFS_PTLRPCD &&
1676             lov_finished_set(set)) {
1677                 lov_statfs_interpret(NULL, set, set->set_count !=
1678                                      cfs_atomic_read(&set->set_success));
1679                 if (lov->lov_qos.lq_statfs_in_progress)
1680                         qos_statfs_done(lov);
1681         }
1682
1683         RETURN(0);
1684 }
1685
1686 int lov_prep_statfs_set(struct obd_device *obd, struct obd_info *oinfo,
1687                         struct lov_request_set **reqset)
1688 {
1689         struct lov_request_set *set;
1690         struct lov_obd *lov = &obd->u.lov;
1691         int rc = 0, i;
1692         ENTRY;
1693
1694         OBD_ALLOC(set, sizeof(*set));
1695         if (set == NULL)
1696                 RETURN(-ENOMEM);
1697         lov_init_set(set);
1698
1699         set->set_obd = obd;
1700         set->set_oi = oinfo;
1701
1702         /* We only get block data from the OBD */
1703         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1704                 struct lov_request *req;
1705
1706                 if (!lov->lov_tgts[i] || (!lov->lov_tgts[i]->ltd_active
1707                                           && (oinfo->oi_flags & OBD_STATFS_NODELAY))) {
1708                         CDEBUG(D_HA, "lov idx %d inactive\n", i);
1709                         continue;
1710                 }
1711
1712                 /* skip targets that have been explicitely disabled by the
1713                  * administrator */
1714                 if (!lov->lov_tgts[i]->ltd_exp) {
1715                         CDEBUG(D_HA, "lov idx %d administratively disabled\n", i);
1716                         continue;
1717                 }
1718
1719                 OBD_ALLOC(req, sizeof(*req));
1720                 if (req == NULL)
1721                         GOTO(out_set, rc = -ENOMEM);
1722
1723                 OBD_ALLOC(req->rq_oi.oi_osfs, sizeof(*req->rq_oi.oi_osfs));
1724                 if (req->rq_oi.oi_osfs == NULL) {
1725                         OBD_FREE(req, sizeof(*req));
1726                         GOTO(out_set, rc = -ENOMEM);
1727                 }
1728
1729                 req->rq_idx = i;
1730                 req->rq_oi.oi_cb_up = cb_statfs_update;
1731                 req->rq_oi.oi_flags = oinfo->oi_flags;
1732
1733                 lov_set_add_req(req, set);
1734         }
1735         if (!set->set_count)
1736                 GOTO(out_set, rc = -EIO);
1737         *reqset = set;
1738         RETURN(rc);
1739 out_set:
1740         lov_fini_statfs_set(set);
1741         RETURN(rc);
1742 }