lustre/lov/lov_obd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  * Author: Phil Schwan <phil@clusterfs.com>
   6  *         Peter Braam <braam@clusterfs.com>
   7  *         Mike Shaver <shaver@clusterfs.com>
   8  *
   9  *   This file is part of Lustre, http://www.lustre.org.
  10  *
  11  *   Lustre is free software; you can redistribute it and/or
  12  *   modify it under the terms of version 2 of the GNU General Public
  13  *   License as published by the Free Software Foundation.
  14  *
  15  *   Lustre is distributed in the hope that it will be useful,
  16  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  *   GNU General Public License for more details.
  19  *
  20  *   You should have received a copy of the GNU General Public License
  21  *   along with Lustre; if not, write to the Free Software
  22  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  23  */
  24
  25 #ifndef EXPORT_SYMTAB
  26 # define EXPORT_SYMTAB
  27 #endif
  28 #define DEBUG_SUBSYSTEM S_LOV
  29 #ifdef __KERNEL__
  30 #include <linux/slab.h>
  31 #include <linux/module.h>
  32 #include <linux/init.h>
  33 #include <linux/slab.h>
  34 #include <linux/pagemap.h>
  35 #include <linux/seq_file.h>
  36 #include <asm/div64.h>
  37 #else
  38 #include <liblustre.h>
  39 #endif
  40
  41 #include <linux/obd_support.h>
  42 #include <linux/lustre_lib.h>
  43 #include <linux/lustre_net.h>
  44 #include <linux/lustre_idl.h>
  45 #include <linux/lustre_dlm.h>
  46 #include <linux/lustre_mds.h>
  47 #include <linux/obd_class.h>
  48 #include <linux/obd_lov.h>
  49 #include <linux/obd_ost.h>
  50 #include <linux/lprocfs_status.h>
  51
  52 #include "lov_internal.h"
  53
  54 static int lov_stripe_offset(struct lov_stripe_md *lsm, obd_off lov_off,
  55                              int stripeno, obd_off *obd_off);
  56
  57 struct lov_lock_handles {
  58         struct portals_handle llh_handle;
  59         atomic_t llh_refcount;
  60         int llh_stripe_count;
  61         struct lustre_handle llh_handles[0];
  62 };
  63
  64 static void lov_llh_addref(void *llhp)
  65 {
  66         struct lov_lock_handles *llh = llhp;
  67
  68         atomic_inc(&llh->llh_refcount);
  69         CDEBUG(D_INFO, "GETting llh %p : new refcount %d\n", llh,
  70                atomic_read(&llh->llh_refcount));
  71 }
  72
  73 static struct lov_lock_handles *lov_llh_new(struct lov_stripe_md *lsm)
  74 {
  75         struct lov_lock_handles *llh;
  76
  77         OBD_ALLOC(llh, sizeof *llh +
  78                   sizeof(*llh->llh_handles) * lsm->lsm_stripe_count);
  79         if (llh == NULL) {
  80                 CERROR("out of memory\n");
  81                 return NULL;
  82         }
  83         atomic_set(&llh->llh_refcount, 2);
  84         llh->llh_stripe_count = lsm->lsm_stripe_count;
  85         INIT_LIST_HEAD(&llh->llh_handle.h_link);
  86         class_handle_hash(&llh->llh_handle, lov_llh_addref);
  87         return llh;
  88 }
  89
  90 static struct lov_lock_handles *lov_handle2llh(struct lustre_handle *handle)
  91 {
  92         ENTRY;
  93         LASSERT(handle != NULL);
  94         RETURN(class_handle2object(handle->cookie));
  95 }
  96
  97 static void lov_llh_put(struct lov_lock_handles *llh)
  98 {
  99         CDEBUG(D_INFO, "PUTting llh %p : new refcount %d\n", llh,
 100                atomic_read(&llh->llh_refcount) - 1);
 101         LASSERT(atomic_read(&llh->llh_refcount) > 0 &&
 102                 atomic_read(&llh->llh_refcount) < 0x5a5a);
 103         if (atomic_dec_and_test(&llh->llh_refcount)) {
 104                 LASSERT(list_empty(&llh->llh_handle.h_link));
 105                 OBD_FREE(llh, sizeof *llh +
 106                          sizeof(*llh->llh_handles) * llh->llh_stripe_count);
 107         }
 108 }
 109
 110 static void lov_llh_destroy(struct lov_lock_handles *llh)
 111 {
 112         class_handle_unhash(&llh->llh_handle);
 113         lov_llh_put(llh);
 114 }
 115
 116 /* obd methods */
 117 static int lov_connect(struct lustre_handle *conn, struct obd_device *obd,
 118                        struct obd_uuid *cluuid)
 119 {
 120         struct ptlrpc_request *req = NULL;
 121         struct lov_obd *lov = &obd->u.lov;
 122         struct lov_desc *desc = &lov->desc;
 123         struct lov_tgt_desc *tgts;
 124         struct obd_export *exp;
 125         int rc, rc2, i;
 126         ENTRY;
 127
 128         rc = class_connect(conn, obd, cluuid);
 129         if (rc)
 130                 RETURN(rc);
 131
 132         exp = class_conn2export(conn);
 133
 134         /* We don't want to actually do the underlying connections more than
 135          * once, so keep track. */
 136         lov->refcount++;
 137         if (lov->refcount > 1) {
 138                 class_export_put(exp);
 139                 RETURN(0);
 140         }
 141
 142         for (i = 0, tgts = lov->tgts; i < desc->ld_tgt_count; i++, tgts++) {
 143                 struct obd_uuid *tgt_uuid = &tgts->uuid;
 144                 struct obd_device *tgt_obd;
 145                 struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
 146                 struct lustre_handle conn = {0, };
 147
 148                 LASSERT( tgt_uuid != NULL);
 149
 150                 tgt_obd = class_find_client_obd(tgt_uuid, LUSTRE_OSC_NAME,
 151                                                 &obd->obd_uuid);
 152
 153                 if (!tgt_obd) {
 154                         CERROR("Target %s not attached\n", tgt_uuid->uuid);
 155                         GOTO(out_disc, rc = -EINVAL);
 156                 }
 157
 158                 if (!tgt_obd->obd_set_up) {
 159                         CERROR("Target %s not set up\n", tgt_uuid->uuid);
 160                         GOTO(out_disc, rc = -EINVAL);
 161                 }
 162
 163                 if (tgt_obd->u.cli.cl_import->imp_invalid) {
 164                         CERROR("not connecting OSC %s; administratively "
 165                                "disabled\n", tgt_uuid->uuid);
 166                         rc = obd_register_observer(tgt_obd, obd);
 167                         if (rc) {
 168                                 CERROR("Target %s register_observer error %d; "
 169                                        "will not be able to reactivate\n",
 170                                        tgt_uuid->uuid, rc);
 171                         }
 172                         continue;
 173                 }
 174
 175                 rc = obd_connect(&conn, tgt_obd, &lov_osc_uuid);
 176                 if (rc) {
 177                         CERROR("Target %s connect error %d\n", tgt_uuid->uuid,
 178                                rc);
 179                         GOTO(out_disc, rc);
 180                 }
 181                 tgts->ltd_exp = class_conn2export(&conn);
 182
 183                 rc = obd_register_observer(tgt_obd, obd);
 184                 if (rc) {
 185                         CERROR("Target %s register_observer error %d\n",
 186                                tgt_uuid->uuid, rc);
 187                         obd_disconnect(tgts->ltd_exp, 0);
 188                         GOTO(out_disc, rc);
 189                 }
 190
 191                 desc->ld_active_tgt_count++;
 192                 tgts->active = 1;
 193         }
 194
 195         ptlrpc_req_finished(req);
 196         class_export_put(exp);
 197         RETURN (0);
 198
 199  out_disc:
 200         while (i-- > 0) {
 201                 struct obd_uuid uuid;
 202                 --tgts;
 203                 --desc->ld_active_tgt_count;
 204                 tgts->active = 0;
 205                 /* save for CERROR below; (we know it's terminated) */
 206                 uuid = tgts->uuid;
 207                 rc2 = obd_disconnect(tgts->ltd_exp, 0);
 208                 if (rc2)
 209                         CERROR("error: LOV target %s disconnect on OST idx %d: "
 210                                "rc = %d\n", uuid.uuid, i, rc2);
 211         }
 212         class_disconnect(exp, 0);
 213         RETURN (rc);
 214 }
 215
 216 static int lov_disconnect(struct obd_export *exp, int flags)
 217 {
 218         struct obd_device *obd = class_exp2obd(exp);
 219         struct lov_obd *lov = &obd->u.lov;
 220         int rc, i;
 221         ENTRY;
 222
 223         if (!lov->tgts)
 224                 goto out_local;
 225
 226         /* Only disconnect the underlying layers on the final disconnect. */
 227         lov->refcount--;
 228         if (lov->refcount != 0)
 229                 goto out_local;
 230
 231         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 232                 if (lov->tgts[i].ltd_exp == NULL)
 233                         continue;
 234
 235                 if (obd->obd_no_recov) {
 236                         /* Pass it on to our clients.
 237                          * XXX This should be an argument to disconnect,
 238                          * XXX not a back-door flag on the OBD.  Ah well.
 239                          */
 240                         struct obd_device *osc_obd;
 241                         osc_obd = class_exp2obd(lov->tgts[i].ltd_exp);
 242                         if (osc_obd)
 243                                 osc_obd->obd_no_recov = 1;
 244                 }
 245
 246                 obd_register_observer(lov->tgts[i].ltd_exp->exp_obd, NULL);
 247
 248                 rc = obd_disconnect(lov->tgts[i].ltd_exp, flags);
 249                 if (rc) {
 250                         if (lov->tgts[i].active) {
 251                                 CERROR("Target %s disconnect error %d\n",
 252                                        lov->tgts[i].uuid.uuid, rc);
 253                         }
 254                         rc = 0;
 255                 }
 256                 if (lov->tgts[i].active) {
 257                         lov->desc.ld_active_tgt_count--;
 258                         lov->tgts[i].active = 0;
 259                 }
 260                 lov->tgts[i].ltd_exp = NULL;
 261         }
 262
 263  out_local:
 264         rc = class_disconnect(exp, 0);
 265         RETURN(rc);
 266 }
 267
 268 /* Error codes:
 269  *
 270  *  -EINVAL  : UUID can't be found in the LOV's target list
 271  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
 272  *  -EBADF   : The UUID is found, but the OBD is the wrong type (!)
 273  */
 274 static int lov_set_osc_active(struct lov_obd *lov, struct obd_uuid *uuid,
 275                               int activate)
 276 {
 277         struct obd_device *obd;
 278         struct lov_tgt_desc *tgt;
 279         int i, rc = 0;
 280         ENTRY;
 281
 282         CDEBUG(D_INFO, "Searching in lov %p for uuid %s (activate=%d)\n",
 283                lov, uuid->uuid, activate);
 284
 285         spin_lock(&lov->lov_lock);
 286         for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) {
 287                 CDEBUG(D_INFO, "lov idx %d is %s conn "LPX64"\n",
 288                        i, tgt->uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
 289                 if (strncmp(uuid->uuid, tgt->uuid.uuid, sizeof uuid->uuid) == 0)
 290                         break;
 291         }
 292
 293         if (i == lov->desc.ld_tgt_count)
 294                 GOTO(out, rc = -EINVAL);
 295
 296         obd = class_exp2obd(tgt->ltd_exp);
 297         if (obd == NULL) {
 298                 /* This can happen if OST failure races with node shutdown */
 299                 GOTO(out, rc = -ENOTCONN);
 300         }
 301
 302         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LOV idx %d\n",
 303                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 304                obd->obd_type->typ_name, i);
 305         LASSERT(strcmp(obd->obd_type->typ_name, "osc") == 0);
 306
 307         if (tgt->active == activate) {
 308                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 309                        activate ? "" : "in");
 310                 GOTO(out, rc);
 311         }
 312
 313         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd, activate ? "" : "in");
 314
 315         tgt->active = activate;
 316         if (activate)
 317                 lov->desc.ld_active_tgt_count++;
 318         else
 319                 lov->desc.ld_active_tgt_count--;
 320
 321         EXIT;
 322  out:
 323         spin_unlock(&lov->lov_lock);
 324         return rc;
 325 }
 326
 327 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
 328                        int active)
 329 {
 330         int rc;
 331         struct obd_uuid *uuid;
 332
 333         if (strcmp(watched->obd_type->typ_name, "osc")) {
 334                 CERROR("unexpected notification of %s %s!\n",
 335                        watched->obd_type->typ_name,
 336                        watched->obd_name);
 337                 return -EINVAL;
 338         }
 339         uuid = &watched->u.cli.cl_import->imp_target_uuid;
 340
 341         /*
 342          * Must notify (MDS) before we mark the OSC as active, so that
 343          * the orphan deletion happens without interference from racing
 344          * creates.
 345          */
 346         if (obd->obd_observer) {
 347                 /* Pass the notification up the chain. */
 348                 rc = obd_notify(obd->obd_observer, watched, active);
 349                 if (rc)
 350                         RETURN(rc);
 351         }
 352
 353         rc = lov_set_osc_active(&obd->u.lov, uuid, active);
 354
 355         if (rc) {
 356                 CERROR("%sactivation of %s failed: %d\n",
 357                        active ? "" : "de", uuid->uuid, rc);
 358         }
 359         RETURN(rc);
 360 }
 361
 362 static int lov_setup(struct obd_device *obd, obd_count len, void *buf)
 363 {
 364         struct lprocfs_static_vars lvars;
 365         struct lustre_cfg *lcfg = buf;
 366         struct lov_desc *desc;
 367         struct lov_obd *lov = &obd->u.lov;
 368         struct obd_uuid *uuids;
 369         struct lov_tgt_desc *tgts;
 370         int i;
 371         int count;
 372         ENTRY;
 373
 374         if (lcfg->lcfg_inllen1 < 1) {
 375                 CERROR("LOV setup requires a descriptor\n");
 376                 RETURN(-EINVAL);
 377         }
 378
 379         if (lcfg->lcfg_inllen2 < 1) {
 380                 CERROR("LOV setup requires an OST UUID list\n");
 381                 RETURN(-EINVAL);
 382         }
 383
 384         desc = (struct lov_desc *)lcfg->lcfg_inlbuf1;
 385         if (sizeof(*desc) > lcfg->lcfg_inllen1) {
 386                 CERROR("descriptor size wrong: %d > %d\n",
 387                        (int)sizeof(*desc), lcfg->lcfg_inllen1);
 388                 RETURN(-EINVAL);
 389         }
 390
 391         count = desc->ld_tgt_count;
 392         uuids = (struct obd_uuid *)lcfg->lcfg_inlbuf2;
 393         if (sizeof(*uuids) * count != lcfg->lcfg_inllen2) {
 394                 CERROR("UUID array size wrong: %u * %u != %u\n",
 395                        (int)sizeof(*uuids), count, lcfg->lcfg_inllen2);
 396                 RETURN(-EINVAL);
 397         }
 398
 399         /* Because of 64-bit divide/mod operations only work with a 32-bit
 400          * divisor in a 32-bit kernel, we cannot support a stripe width
 401          * of 4GB or larger on 32-bit CPUs.
 402          */
 403         if ((desc->ld_default_stripe_count ?
 404              desc->ld_default_stripe_count : desc->ld_tgt_count) *
 405              desc->ld_default_stripe_size > ~0UL) {
 406                 CERROR("LOV: stripe width "LPU64"x%u > %lu on 32-bit system\n",
 407                        desc->ld_default_stripe_size,
 408                        desc->ld_default_stripe_count ?
 409                        desc->ld_default_stripe_count : desc->ld_tgt_count,~0UL);
 410                 RETURN(-EINVAL);
 411         }
 412
 413         lov->bufsize = sizeof(struct lov_tgt_desc) * count;
 414         OBD_ALLOC(lov->tgts, lov->bufsize);
 415         if (lov->tgts == NULL) {
 416                 CERROR("Out of memory\n");
 417                 RETURN(-EINVAL);
 418         }
 419
 420         lov->desc = *desc;
 421         spin_lock_init(&lov->lov_lock);
 422
 423         for (i = 0, tgts = lov->tgts; i < desc->ld_tgt_count; i++, tgts++) {
 424                 struct obd_uuid *uuid = &tgts->uuid;
 425
 426                 /* NULL termination already checked */
 427                 *uuid = uuids[i];
 428         }
 429
 430         lprocfs_init_vars(lov, &lvars);
 431         lprocfs_obd_setup(obd, lvars.obd_vars);
 432 #ifdef __KERNEL__
 433         {
 434                 struct proc_dir_entry *entry;
 435
 436                 entry = create_proc_entry("target_obd", 0444,
 437                                           obd->obd_proc_entry);
 438                 if (entry != NULL) {
 439                         entry->proc_fops = &lov_proc_target_fops;
 440                         entry->data = obd;
 441                 }
 442         }
 443 #endif
 444
 445         RETURN(0);
 446 }
 447
 448 static int lov_cleanup(struct obd_device *obd, int flags)
 449 {
 450         struct lov_obd *lov = &obd->u.lov;
 451
 452         lprocfs_obd_cleanup(obd);
 453         OBD_FREE(lov->tgts, lov->bufsize);
 454
 455         RETURN(0);
 456 }
 457
 458
 459 /* compute object size given "stripeno" and the ost size */
 460 static obd_size lov_stripe_size(struct lov_stripe_md *lsm, obd_size ost_size,
 461                                 int stripeno)
 462 {
 463         unsigned long ssize  = lsm->lsm_stripe_size;
 464         unsigned long swidth = ssize * lsm->lsm_stripe_count;
 465         unsigned long stripe_size;
 466         obd_size lov_size;
 467
 468         if (ost_size == 0)
 469                 return 0;
 470
 471         /* do_div(a, b) returns a % b, and a = a / b */
 472         stripe_size = do_div(ost_size, ssize);
 473         if (stripe_size)
 474                 lov_size = ost_size * swidth + stripeno * ssize + stripe_size;
 475         else
 476                 lov_size = (ost_size - 1) * swidth + (stripeno + 1) * ssize;
 477
 478         return lov_size;
 479 }
 480
 481 static void lov_merge_attrs(struct obdo *tgt, struct obdo *src, obd_flag valid,
 482                             struct lov_stripe_md *lsm, int stripeno, int *set)
 483 {
 484         valid &= src->o_valid;
 485
 486         if (*set) {
 487                 if (valid & OBD_MD_FLSIZE) {
 488                         /* this handles sparse files properly */
 489                         obd_size lov_size;
 490
 491                         lov_size = lov_stripe_size(lsm, src->o_size, stripeno);
 492                         if (lov_size > tgt->o_size)
 493                                 tgt->o_size = lov_size;
 494                 }
 495                 if (valid & OBD_MD_FLBLOCKS)
 496                         tgt->o_blocks += src->o_blocks;
 497                 if (valid & OBD_MD_FLBLKSZ)
 498                         tgt->o_blksize += src->o_blksize;
 499                 if (valid & OBD_MD_FLCTIME && tgt->o_ctime < src->o_ctime)
 500                         tgt->o_ctime = src->o_ctime;
 501                 if (valid & OBD_MD_FLMTIME && tgt->o_mtime < src->o_mtime)
 502                         tgt->o_mtime = src->o_mtime;
 503         } else {
 504                 memcpy(tgt, src, sizeof(*tgt));
 505                 tgt->o_id = lsm->lsm_object_id;
 506                 if (valid & OBD_MD_FLSIZE)
 507                         tgt->o_size = lov_stripe_size(lsm,src->o_size,stripeno);
 508                 *set = 1;
 509         }
 510 }
 511
 512 #ifndef log2
 513 #define log2(n) ffz(~(n))
 514 #endif
 515
 516 static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa,
 517                              struct lov_stripe_md **ea,
 518                              struct obd_trans_info *oti)
 519 {
 520         struct lov_obd *lov;
 521         struct obdo *tmp_oa;
 522         struct obd_uuid *ost_uuid = NULL;
 523         int rc = 0, i;
 524         ENTRY;
 525
 526         LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS &&
 527                 src_oa->o_flags == OBD_FL_DELORPHAN);
 528
 529         lov = &export->exp_obd->u.lov;
 530
 531         tmp_oa = obdo_alloc();
 532         if (tmp_oa == NULL)
 533                 RETURN(-ENOMEM);
 534
 535         if (src_oa->o_valid & OBD_MD_FLINLINE) {
 536                 ost_uuid = (struct obd_uuid *)src_oa->o_inline;
 537                 CDEBUG(D_HA, "clearing orphans only for %s\n",
 538                        ost_uuid->uuid);
 539         }
 540
 541         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 542                 struct lov_stripe_md obj_md;
 543                 struct lov_stripe_md *obj_mdp = &obj_md;
 544                 int err;
 545
 546                 /* if called for a specific target, we don't
 547                    care if it is not active. */
 548                 if (lov->tgts[i].active == 0 && ost_uuid == NULL) {
 549                         CDEBUG(D_HA, "lov idx %d inactive\n", i);
 550                         continue;
 551                 }
 552
 553                 if (ost_uuid && !obd_uuid_equals(ost_uuid, &lov->tgts[i].uuid))
 554                         continue;
 555
 556                 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
 557
 558                 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
 559                 err = obd_create(lov->tgts[i].ltd_exp, tmp_oa, &obj_mdp, oti);
 560                 if (err)
 561                         /* This export will be disabled until it is recovered,
 562                            and then orphan recovery will be completed. */
 563                         CERROR("error in orphan recovery on OST idx %d/%d: "
 564                                "rc = %d\n", i, lov->desc.ld_tgt_count, err);
 565
 566                 if (ost_uuid)
 567                         break;
 568         }
 569         obdo_free(tmp_oa);
 570         RETURN(rc);
 571 }
 572
 573 #define LOV_CREATE_RESEED_INTERVAL 1000
 574
 575 /* the LOV expects oa->o_id to be set to the LOV object id */
 576 static int lov_create(struct obd_export *exp, struct obdo *src_oa,
 577                       struct lov_stripe_md **ea, struct obd_trans_info *oti)
 578 {
 579         static int ost_start_idx, ost_start_count;
 580         struct lov_obd *lov;
 581         struct lov_stripe_md *lsm;
 582         struct lov_oinfo *loi = NULL;
 583         struct obdo *tmp_oa, *ret_oa;
 584         struct llog_cookie *cookies = NULL;
 585         unsigned ost_count, ost_idx;
 586         int set = 0, obj_alloc = 0, cookie_sent = 0, rc = 0, i;
 587         ENTRY;
 588
 589         LASSERT(ea != NULL);
 590
 591         if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
 592             src_oa->o_flags == OBD_FL_DELORPHAN) {
 593                 rc = lov_clear_orphans(exp, src_oa, ea, oti);
 594                 RETURN(rc);
 595         }
 596
 597         if (exp == NULL)
 598                 RETURN(-EINVAL);
 599
 600         lov = &exp->exp_obd->u.lov;
 601
 602         if (!lov->desc.ld_active_tgt_count)
 603                 RETURN(-EIO);
 604
 605         /* Recreate a specific object id at the given OST index */
 606         if (src_oa->o_valid & OBD_MD_FLFLAGS && src_oa->o_flags &
 607                                                 OBD_FL_RECREATE_OBJS) {
 608                  struct lov_stripe_md obj_md;
 609                  struct lov_stripe_md *obj_mdp = &obj_md;
 610
 611                  ost_idx = src_oa->o_nlink;
 612                  lsm = *ea;
 613                  if (lsm == NULL)
 614                         RETURN(-EINVAL);
 615                  if (ost_idx >= lov->desc.ld_tgt_count)
 616                          RETURN(-EINVAL);
 617                  for (i = 0; i < lsm->lsm_stripe_count; i++) {
 618                          if (lsm->lsm_oinfo[i].loi_ost_idx == ost_idx) {
 619                                  if (lsm->lsm_oinfo[i].loi_id != src_oa->o_id)
 620                                          RETURN(-EINVAL);
 621                                  break;
 622                          }
 623                  }
 624                  if (i == lsm->lsm_stripe_count)
 625                          RETURN(-EINVAL);
 626
 627                  rc = obd_create(lov->tgts[ost_idx].ltd_exp, src_oa,
 628                                  &obj_mdp, oti);
 629                  RETURN(rc);
 630         }
 631
 632         ret_oa = obdo_alloc();
 633         if (!ret_oa)
 634                 RETURN(-ENOMEM);
 635
 636         tmp_oa = obdo_alloc();
 637         if (!tmp_oa)
 638                 GOTO(out_oa, rc = -ENOMEM);
 639
 640         lsm = *ea;
 641         if (lsm == NULL) {
 642                 int stripes;
 643                 ost_count = lov_get_stripecnt(lov, 0);
 644
 645                 /* If the MDS file was truncated up to some size, stripe over
 646                  * enough OSTs to allow the file to be created at that size. */
 647                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 648                         stripes=((src_oa->o_size+LUSTRE_STRIPE_MAXBYTES)>>12)-1;
 649                         do_div(stripes, (__u32)(LUSTRE_STRIPE_MAXBYTES >> 12));
 650
 651                         if (stripes > lov->desc.ld_active_tgt_count)
 652                                 RETURN(-EFBIG);
 653                         if (stripes < ost_count)
 654                                 stripes = ost_count;
 655                 } else {
 656                         stripes = ost_count;
 657                 }
 658
 659                 rc = lov_alloc_memmd(&lsm, stripes, lov->desc.ld_pattern ?
 660                                      lov->desc.ld_pattern : LOV_PATTERN_RAID0);
 661                 if (rc < 0)
 662                         GOTO(out_tmp, rc);
 663
 664                 rc = 0;
 665         }
 666
 667         ost_count = lov->desc.ld_tgt_count;
 668
 669         LASSERT(src_oa->o_valid & OBD_MD_FLID);
 670         lsm->lsm_object_id = src_oa->o_id;
 671         if (!lsm->lsm_stripe_size)
 672                 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
 673         if (!lsm->lsm_pattern) {
 674                 lsm->lsm_pattern = lov->desc.ld_pattern ?
 675                         lov->desc.ld_pattern : LOV_PATTERN_RAID0;
 676         }
 677
 678         if (*ea == NULL || lsm->lsm_oinfo[0].loi_ost_idx >= ost_count) {
 679                 if (--ost_start_count <= 0) {
 680                         ost_start_idx = ll_insecure_random_int();
 681                         ost_start_count = LOV_CREATE_RESEED_INTERVAL;
 682                 } else if (lsm->lsm_stripe_count >=
 683                            lov->desc.ld_active_tgt_count) {
 684                         /* If we allocate from all of the stripes, make the
 685                          * next file start on the next OST. */
 686                         ++ost_start_idx;
 687                 }
 688                 ost_idx = ost_start_idx % ost_count;
 689         } else {
 690                 ost_idx = lsm->lsm_oinfo[0].loi_ost_idx;
 691         }
 692
 693         CDEBUG(D_INODE, "allocating %d subobjs for objid "LPX64" at idx %d\n",
 694                lsm->lsm_stripe_count, lsm->lsm_object_id, ost_idx);
 695
 696         /* XXX LOV STACKING: need to figure out how many real OSCs */
 697         if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
 698                 oti_alloc_cookies(oti, lsm->lsm_stripe_count);
 699                 if (!oti->oti_logcookies)
 700                         GOTO(out_cleanup, rc = -ENOMEM);
 701                 cookies = oti->oti_logcookies;
 702         }
 703
 704         loi = lsm->lsm_oinfo;
 705         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 706                 struct lov_stripe_md obj_md;
 707                 struct lov_stripe_md *obj_mdp = &obj_md;
 708                 int err;
 709
 710                 ++ost_start_idx;
 711                 if (lov->tgts[ost_idx].active == 0) {
 712                         CDEBUG(D_HA, "lov idx %d inactive\n", ost_idx);
 713                         continue;
 714                 }
 715
 716                 /* create data objects with "parent" OA */
 717                 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
 718
 719                 /* XXX When we start creating objects on demand, we need to
 720                  *     make sure that we always create the object on the
 721                  *     stripe which holds the existing file size.
 722                  */
 723                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 724                         if (lov_stripe_offset(lsm, src_oa->o_size, i,
 725                                               &tmp_oa->o_size) < 0 &&
 726                             tmp_oa->o_size)
 727                                 tmp_oa->o_size--;
 728
 729                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
 730                                i, tmp_oa->o_size, src_oa->o_size);
 731                 }
 732
 733
 734                 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
 735                 err = obd_create(lov->tgts[ost_idx].ltd_exp, tmp_oa, &obj_mdp,
 736                                  oti);
 737                 if (err) {
 738                         if (lov->tgts[ost_idx].active) {
 739                                 CERROR("error creating objid "LPX64" sub-object"
 740                                        " on OST idx %d/%d: rc = %d\n",
 741                                        src_oa->o_id, ost_idx,
 742                                        lsm->lsm_stripe_count, err);
 743                                 if (err > 0) {
 744                                         CERROR("obd_create returned invalid "
 745                                                "err %d\n", err);
 746                                         err = -EIO;
 747                                 }
 748                         }
 749                         if (!rc)
 750                                 rc = err;
 751                         continue;
 752                 }
 753                 if (oti->oti_objid)
 754                         oti->oti_objid[ost_idx] = tmp_oa->o_id;
 755                 loi->loi_id = tmp_oa->o_id;
 756                 loi->loi_ost_idx = ost_idx;
 757                 CDEBUG(D_INODE, "objid "LPX64" has subobj "LPX64" at idx %d\n",
 758                        lsm->lsm_object_id, loi->loi_id, ost_idx);
 759
 760                 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
 761                                 obj_alloc, &set);
 762                 loi_init(loi);
 763
 764                 if (cookies)
 765                         ++oti->oti_logcookies;
 766                 if (tmp_oa->o_valid & OBD_MD_FLCOOKIE)
 767                         ++cookie_sent;
 768                 ++obj_alloc;
 769                 ++loi;
 770
 771                 /* If we have allocated enough objects, we are OK */
 772                 if (obj_alloc == lsm->lsm_stripe_count)
 773                         GOTO(out_done, rc = 0);
 774         }
 775
 776         if (obj_alloc == 0) {
 777                 if (rc == 0)
 778                         rc = -EIO;
 779                 GOTO(out_cleanup, rc);
 780         }
 781
 782         /* If we were passed specific striping params, then a failure to
 783          * meet those requirements is an error, since we can't reallocate
 784          * that memory (it might be part of a larger array or something).
 785          *
 786          * We can only get here if lsm_stripe_count was originally > 1.
 787          */
 788         if (*ea != NULL) {
 789                 CERROR("can't lstripe objid "LPX64": have %u want %u, rc %d\n",
 790                        lsm->lsm_object_id, obj_alloc, lsm->lsm_stripe_count,rc);
 791                 if (rc == 0)
 792                         rc = -EFBIG;
 793                 GOTO(out_cleanup, rc);
 794         } else {
 795                 struct lov_stripe_md *lsm_new;
 796                 /* XXX LOV STACKING call into osc for sizes */
 797                 unsigned oldsize, newsize;
 798
 799                 if (oti && cookies && cookie_sent) {
 800                         oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
 801                         newsize = obj_alloc * sizeof(*cookies);
 802
 803                         oti_alloc_cookies(oti, obj_alloc);
 804                         if (oti->oti_logcookies) {
 805                                 memcpy(oti->oti_logcookies, cookies, newsize);
 806                                 OBD_FREE(cookies, oldsize);
 807                                 cookies = oti->oti_logcookies;
 808                         } else {
 809                                 CWARN("'leaking' %d bytes\n", oldsize-newsize);
 810                         }
 811                 }
 812
 813                 CWARN("using fewer stripes for object "LPX64": old %u new %u\n",
 814                       lsm->lsm_object_id, lsm->lsm_stripe_count, obj_alloc);
 815                 oldsize = lov_stripe_md_size(lsm->lsm_stripe_count);
 816                 newsize = lov_stripe_md_size(obj_alloc);
 817                 OBD_ALLOC(lsm_new, newsize);
 818                 if (lsm_new != NULL) {
 819                         memcpy(lsm_new, lsm, newsize);
 820                         lsm_new->lsm_stripe_count = obj_alloc;
 821                         OBD_FREE(lsm, oldsize);
 822                         lsm = lsm_new;
 823                 } else {
 824                         CWARN("'leaking' %d bytes\n", oldsize - newsize);
 825                 }
 826                 rc = 0;
 827         }
 828         EXIT;
 829  out_done:
 830         *ea = lsm;
 831         if (src_oa->o_valid & OBD_MD_FLSIZE &&
 832             ret_oa->o_size != src_oa->o_size) {
 833                 CERROR("original size "LPU64" isn't new object size "LPU64"\n",
 834                        src_oa->o_size, ret_oa->o_size);
 835                 LBUG();
 836         }
 837         ret_oa->o_id = src_oa->o_id;
 838         memcpy(src_oa, ret_oa, sizeof(*src_oa));
 839
 840  out_tmp:
 841         obdo_free(tmp_oa);
 842  out_oa:
 843         obdo_free(ret_oa);
 844         if (oti && cookies) {
 845                 oti->oti_logcookies = cookies;
 846                 if (!cookie_sent) {
 847                         oti_free_cookies(oti);
 848                         src_oa->o_valid &= ~OBD_MD_FLCOOKIE;
 849                 } else {
 850                         src_oa->o_valid |= OBD_MD_FLCOOKIE;
 851                 }
 852         }
 853         RETURN(rc);
 854
 855  out_cleanup:
 856         while (obj_alloc-- > 0) {
 857                 struct obd_export *sub_exp;
 858                 int err;
 859
 860                 --loi;
 861                 sub_exp = lov->tgts[loi->loi_ost_idx].ltd_exp;
 862                 /* destroy already created objects here */
 863                 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
 864                 tmp_oa->o_id = loi->loi_id;
 865
 866                 err = obd_destroy(sub_exp, tmp_oa, NULL, oti);
 867                 if (err)
 868                         CERROR("Failed to uncreate objid "LPX64" subobj "LPX64
 869                                " on OST idx %d: rc = %d\n", src_oa->o_id,
 870                                loi->loi_id, loi->loi_ost_idx, err);
 871         }
 872         if (*ea == NULL)
 873                 obd_free_memmd(exp, &lsm);
 874         goto out_tmp;
 875 }
 876
 877 #define lsm_bad_magic(LSMP)                                     \
 878 ({                                                              \
 879         struct lov_stripe_md *_lsm__ = (LSMP);                  \
 880         int _ret__ = 0;                                         \
 881         if (!_lsm__) {                                          \
 882                 CERROR("LOV requires striping ea\n");           \
 883                 _ret__ = 1;                                     \
 884         } else if (_lsm__->lsm_magic != LOV_MAGIC) {            \
 885                 CERROR("LOV striping magic bad %#x != %#x\n",   \
 886                        _lsm__->lsm_magic, LOV_MAGIC);           \
 887                 _ret__ = 1;                                     \
 888         }                                                       \
 889         _ret__;                                                 \
 890 })
 891
 892 static int lov_destroy(struct obd_export *exp, struct obdo *oa,
 893                        struct lov_stripe_md *lsm, struct obd_trans_info *oti)
 894 {
 895         struct obdo tmp;
 896         struct lov_obd *lov;
 897         struct lov_oinfo *loi;
 898         int rc = 0, i;
 899         ENTRY;
 900
 901         if (lsm_bad_magic(lsm))
 902                 RETURN(-EINVAL);
 903
 904         if (!exp || !exp->exp_obd)
 905                 RETURN(-ENODEV);
 906
 907         lov = &exp->exp_obd->u.lov;
 908         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
 909                 int err;
 910                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
 911                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 912                         /* Orphan clean up will (someday) fix this up. */
 913                         if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE)
 914                                 oti->oti_logcookies++;
 915                         continue;
 916                 }
 917
 918                 memcpy(&tmp, oa, sizeof(tmp));
 919                 tmp.o_id = loi->loi_id;
 920                 err = obd_destroy(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
 921                                   NULL, oti);
 922                 if (err && lov->tgts[loi->loi_ost_idx].active) {
 923                         CERROR("error: destroying objid "LPX64" subobj "
 924                                LPX64" on OST idx %d: rc = %d\n",
 925                                oa->o_id, loi->loi_id, loi->loi_ost_idx, err);
 926                         if (!rc)
 927                                 rc = err;
 928                 }
 929         }
 930         RETURN(rc);
 931 }
 932
 933 static int lov_getattr(struct obd_export *exp, struct obdo *oa,
 934                        struct lov_stripe_md *lsm)
 935 {
 936         struct obdo tmp;
 937         struct lov_obd *lov;
 938         struct lov_oinfo *loi;
 939         int i, rc = 0, set = 0;
 940         ENTRY;
 941
 942         if (lsm_bad_magic(lsm))
 943                 RETURN(-EINVAL);
 944
 945         if (!exp || !exp->exp_obd)
 946                 RETURN(-ENODEV);
 947
 948         lov = &exp->exp_obd->u.lov;
 949
 950         CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
 951                lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
 952         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
 953                 int err;
 954
 955                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
 956                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 957                         continue;
 958                 }
 959
 960                 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
 961                        "%u\n", oa->o_id, i, loi->loi_id, loi->loi_ost_idx);
 962                 /* create data objects with "parent" OA */
 963                 memcpy(&tmp, oa, sizeof(tmp));
 964                 tmp.o_id = loi->loi_id;
 965
 966                 err = obd_getattr(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
 967                                   NULL);
 968                 if (err) {
 969                         if (lov->tgts[loi->loi_ost_idx].active) {
 970                                 CERROR("error: getattr objid "LPX64" subobj "
 971                                        LPX64" on OST idx %d: rc = %d\n",
 972                                        oa->o_id, loi->loi_id, loi->loi_ost_idx,
 973                                        err);
 974                                 RETURN(err);
 975                         }
 976                 } else {
 977                         lov_merge_attrs(oa, &tmp, tmp.o_valid, lsm, i, &set);
 978                 }
 979         }
 980         if (!set)
 981                 rc = -EIO;
 982         RETURN(rc);
 983 }
 984
 985 static int lov_getattr_interpret(struct ptlrpc_request_set *rqset, void *data,
 986                                  int rc)
 987 {
 988         struct lov_getattr_async_args *aa = data;
 989         struct lov_stripe_md *lsm = aa->aa_lsm;
 990         struct obdo          *oa = aa->aa_oa;
 991         struct obdo          *obdos = aa->aa_obdos;
 992         struct lov_oinfo     *loi;
 993         int                   i;
 994         int                   set = 0;
 995         ENTRY;
 996
 997         if (rc == 0) {
 998                 /* NB all stripe requests succeeded to get here */
 999
1000                 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1001                      i++, loi++) {
1002                         if (obdos[i].o_valid == 0)      /* inactive stripe */
1003                                 continue;
1004
1005                         lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1006                                         i, &set);
1007                 }
1008
1009                 if (!set) {
1010                         CERROR ("No stripes had valid attrs\n");
1011                         rc = -EIO;
1012                 }
1013         }
1014
1015         OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1016         RETURN (rc);
1017 }
1018
1019 static int lov_getattr_async(struct obd_export *exp, struct obdo *oa,
1020                               struct lov_stripe_md *lsm,
1021                               struct ptlrpc_request_set *rqset)
1022 {
1023         struct obdo *obdos;
1024         struct lov_obd *lov;
1025         struct lov_oinfo *loi;
1026         struct lov_getattr_async_args *aa;
1027         int i, rc = 0, set = 0;
1028         ENTRY;
1029
1030         if (lsm_bad_magic(lsm))
1031                 RETURN(-EINVAL);
1032
1033         if (!exp || !exp->exp_obd)
1034                 RETURN(-ENODEV);
1035
1036         lov = &exp->exp_obd->u.lov;
1037
1038         OBD_ALLOC (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1039         if (obdos == NULL)
1040                 RETURN(-ENOMEM);
1041
1042         CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1043                lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
1044         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1045                 int err;
1046
1047                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1048                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1049                         /* leaves obdos[i].obd_valid unset */
1050                         continue;
1051                 }
1052
1053                 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
1054                        "%u\n", oa->o_id, i, loi->loi_id, loi->loi_ost_idx);
1055                 /* create data objects with "parent" OA */
1056                 memcpy(&obdos[i], oa, sizeof(obdos[i]));
1057                 obdos[i].o_id = loi->loi_id;
1058
1059                 err = obd_getattr_async(lov->tgts[loi->loi_ost_idx].ltd_exp,
1060                                          &obdos[i], NULL, rqset);
1061                 if (err) {
1062                         CERROR("error: getattr objid "LPX64" subobj "
1063                                LPX64" on OST idx %d: rc = %d\n",
1064                                oa->o_id, loi->loi_id, loi->loi_ost_idx,
1065                                err);
1066                         GOTO(out_obdos, rc = err);
1067                 }
1068                 set = 1;
1069         }
1070         if (!set)
1071                 GOTO (out_obdos, rc = -EIO);
1072
1073         LASSERT (rqset->set_interpret == NULL);
1074         rqset->set_interpret = lov_getattr_interpret;
1075         LASSERT (sizeof (rqset->set_args) >= sizeof (*aa));
1076         aa = (struct lov_getattr_async_args *)&rqset->set_args;
1077         aa->aa_lsm = lsm;
1078         aa->aa_oa = oa;
1079         aa->aa_obdos = obdos;
1080         aa->aa_lov = lov;
1081         GOTO(out, rc = 0);
1082
1083 out_obdos:
1084         OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1085 out:
1086         RETURN(rc);
1087 }
1088
1089
1090 static int lov_setattr(struct obd_export *exp, struct obdo *src_oa,
1091                        struct lov_stripe_md *lsm, struct obd_trans_info *oti)
1092 {
1093         struct obdo *tmp_oa, *ret_oa;
1094         struct lov_obd *lov;
1095         struct lov_oinfo *loi;
1096         int rc = 0, i, set = 0;
1097         ENTRY;
1098
1099         if (lsm_bad_magic(lsm))
1100                 RETURN(-EINVAL);
1101
1102         if (!exp || !exp->exp_obd)
1103                 RETURN(-ENODEV);
1104
1105         /* for now, we only expect time updates here */
1106         LASSERT(!(src_oa->o_valid & ~(OBD_MD_FLID|OBD_MD_FLTYPE | OBD_MD_FLMODE|
1107                                       OBD_MD_FLATIME | OBD_MD_FLMTIME |
1108                                       OBD_MD_FLCTIME | OBD_MD_FLFLAGS |
1109                                       OBD_MD_FLSIZE)));
1110         ret_oa = obdo_alloc();
1111         if (!ret_oa)
1112                 RETURN(-ENOMEM);
1113
1114         tmp_oa = obdo_alloc();
1115         if (!tmp_oa)
1116                 GOTO(out_oa, rc = -ENOMEM);
1117
1118         lov = &exp->exp_obd->u.lov;
1119         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1120                 int err;
1121
1122                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1123                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1124                         continue;
1125                 }
1126
1127                 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1128                 tmp_oa->o_id = loi->loi_id;
1129
1130                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1131                         if (lov_stripe_offset(lsm, src_oa->o_size, i,
1132                                               &tmp_oa->o_size) < 0 &&
1133                             tmp_oa->o_size)
1134                                 tmp_oa->o_size--;
1135
1136                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1137                                i, tmp_oa->o_size, src_oa->o_size);
1138                 }
1139
1140                 err = obd_setattr(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp_oa,
1141                                   NULL, NULL);
1142                 if (err) {
1143                         if (lov->tgts[loi->loi_ost_idx].active) {
1144                                 CERROR("error: setattr objid "LPX64" subobj "
1145                                        LPX64" on OST idx %d: rc = %d\n",
1146                                        src_oa->o_id, loi->loi_id,
1147                                        loi->loi_ost_idx, err);
1148                                 if (!rc)
1149                                         rc = err;
1150                         }
1151                         continue;
1152                 }
1153                 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm, i, &set);
1154         }
1155         if (!set && !rc)
1156                 rc = -EIO;
1157
1158         ret_oa->o_id = src_oa->o_id;
1159         memcpy(src_oa, ret_oa, sizeof(*src_oa));
1160         GOTO(out_tmp, rc);
1161 out_tmp:
1162         obdo_free(tmp_oa);
1163 out_oa:
1164         obdo_free(ret_oa);
1165         return rc;
1166 }
1167
1168 /* we have an offset in file backed by an lov and want to find out where
1169  * that offset lands in our given stripe of the file.  for the easy
1170  * case where the offset is within the stripe, we just have to scale the
1171  * offset down to make it relative to the stripe instead of the lov.
1172  *
1173  * the harder case is what to do when the offset doesn't intersect the
1174  * stripe.  callers will want start offsets clamped ahead to the start
1175  * of the nearest stripe in the file.  end offsets similarly clamped to the
1176  * nearest ending byte of a stripe in the file:
1177  *
1178  * all this function does is move offsets to the nearest region of the
1179  * stripe, and it does its work "mod" the full length of all the stripes.
1180  * consider a file with 3 stripes:
1181  *
1182  *             S                                              E
1183  * ---------------------------------------------------------------------
1184  * |    0    |     1     |     2     |    0    |     1     |     2     |
1185  * ---------------------------------------------------------------------
1186  *
1187  * to find stripe 1's offsets for S and E, it divides by the full stripe
1188  * width and does its math in the context of a single set of stripes:
1189  *
1190  *             S         E
1191  * -----------------------------------
1192  * |    0    |     1     |     2     |
1193  * -----------------------------------
1194  *
1195  * it'll notice that E is outside stripe 1 and clamp it to the end of the
1196  * stripe, then multiply it back out by lov_off to give the real offsets in
1197  * the stripe:
1198  *
1199  *   S                   E
1200  * ---------------------------------------------------------------------
1201  * |    1    |     1     |     1     |    1    |     1     |     1     |
1202  * ---------------------------------------------------------------------
1203  *
1204  * it would have done similarly and pulled S forward to the start of a 1
1205  * stripe if, say, S had landed in a 0 stripe.
1206  *
1207  * this rounding isn't always correct.  consider an E lov offset that lands
1208  * on a 0 stripe, the "mod stripe width" math will pull it forward to the
1209  * start of a 1 stripe, when in fact it wanted to be rounded back to the end
1210  * of a previous 1 stripe.  this logic is handled by callers and this is why:
1211  *
1212  * this function returns < 0 when the offset was "before" the stripe and
1213  * was moved forward to the start of the stripe in question;  0 when it
1214  * falls in the stripe and no shifting was done; > 0 when the offset
1215  * was outside the stripe and was pulled back to its final byte. */
1216 static int lov_stripe_offset(struct lov_stripe_md *lsm, obd_off lov_off,
1217                              int stripeno, obd_off *obd_off)
1218 {
1219         unsigned long ssize  = lsm->lsm_stripe_size;
1220         unsigned long swidth = ssize * lsm->lsm_stripe_count;
1221         unsigned long stripe_off, this_stripe;
1222         int ret = 0;
1223
1224         if (lov_off == OBD_OBJECT_EOF) {
1225                 *obd_off = OBD_OBJECT_EOF;
1226                 return 0;
1227         }
1228
1229         /* do_div(a, b) returns a % b, and a = a / b */
1230         stripe_off = do_div(lov_off, swidth);
1231
1232         this_stripe = stripeno * ssize;
1233         if (stripe_off < this_stripe) {
1234                 stripe_off = 0;
1235                 ret = -1;
1236         } else {
1237                 stripe_off -= this_stripe;
1238
1239                 if (stripe_off >= ssize) {
1240                         stripe_off = ssize;
1241                         ret = 1;
1242                 }
1243         }
1244
1245         *obd_off = lov_off * ssize + stripe_off;
1246         return ret;
1247 }
1248
1249 /* Given a whole-file size and a stripe number, give the file size which
1250  * corresponds to the individual object of that stripe.
1251  *
1252  * This behaves basically in the same was as lov_stripe_offset, except that
1253  * file sizes falling before the beginning of a stripe are clamped to the end
1254  * of the previous stripe, not the beginning of the next:
1255  *
1256  *                                               S
1257  * ---------------------------------------------------------------------
1258  * |    0    |     1     |     2     |    0    |     1     |     2     |
1259  * ---------------------------------------------------------------------
1260  *
1261  * if clamped to stripe 2 becomes:
1262  *
1263  *                                   S
1264  * ---------------------------------------------------------------------
1265  * |    0    |     1     |     2     |    0    |     1     |     2     |
1266  * ---------------------------------------------------------------------
1267  */
1268 static obd_off lov_size_to_stripe(struct lov_stripe_md *lsm, obd_off file_size,
1269                                   int stripeno)
1270 {
1271         unsigned long ssize  = lsm->lsm_stripe_size;
1272         unsigned long swidth = ssize * lsm->lsm_stripe_count;
1273         unsigned long stripe_off, this_stripe;
1274
1275         if (file_size == OBD_OBJECT_EOF)
1276                 return OBD_OBJECT_EOF;
1277
1278         /* do_div(a, b) returns a % b, and a = a / b */
1279         stripe_off = do_div(file_size, swidth);
1280
1281         this_stripe = stripeno * ssize;
1282         if (stripe_off < this_stripe) {
1283                 /* Move to end of previous stripe, or zero */
1284                 if (file_size > 0) {
1285                         file_size--;
1286                         stripe_off = ssize;
1287                 } else {
1288                         stripe_off = 0;
1289                 }
1290         } else {
1291                 stripe_off -= this_stripe;
1292
1293                 if (stripe_off >= ssize) {
1294                         /* Clamp to end of this stripe */
1295                         stripe_off = ssize;
1296                 }
1297         }
1298
1299         return (file_size * ssize + stripe_off);
1300 }
1301
1302 /* given an extent in an lov and a stripe, calculate the extent of the stripe
1303  * that is contained within the lov extent.  this returns true if the given
1304  * stripe does intersect with the lov extent. */
1305 static int lov_stripe_intersects(struct lov_stripe_md *lsm, int stripeno,
1306                                  obd_off start, obd_off end,
1307                                  obd_off *obd_start, obd_off *obd_end)
1308 {
1309         int start_side, end_side;
1310
1311         start_side = lov_stripe_offset(lsm, start, stripeno, obd_start);
1312         end_side = lov_stripe_offset(lsm, end, stripeno, obd_end);
1313
1314         CDEBUG(D_INODE, "["LPU64"->"LPU64"] -> [(%d) "LPU64"->"LPU64" (%d)]\n",
1315                start, end, start_side, *obd_start, *obd_end, end_side);
1316
1317         /* this stripe doesn't intersect the file extent when neither
1318          * start or the end intersected the stripe and obd_start and
1319          * obd_end got rounded up to the save value. */
1320         if (start_side != 0 && end_side != 0 && *obd_start == *obd_end)
1321                 return 0;
1322
1323         /* as mentioned in the lov_stripe_offset commentary, end
1324          * might have been shifted in the wrong direction.  This
1325          * happens when an end offset is before the stripe when viewed
1326          * through the "mod stripe size" math. we detect it being shifted
1327          * in the wrong direction and touch it up.
1328          * interestingly, this can't underflow since end must be > start
1329          * if we passed through the previous check.
1330          * (should we assert for that somewhere?) */
1331         if (end_side != 0)
1332                 (*obd_end)--;
1333
1334         return 1;
1335 }
1336
1337 /* compute which stripe number "lov_off" will be written into */
1338 static int lov_stripe_number(struct lov_stripe_md *lsm, obd_off lov_off)
1339 {
1340         unsigned long ssize  = lsm->lsm_stripe_size;
1341         unsigned long swidth = ssize * lsm->lsm_stripe_count;
1342         unsigned long stripe_off;
1343
1344         stripe_off = do_div(lov_off, swidth);
1345
1346         return stripe_off / ssize;
1347 }
1348
1349 /* FIXME: maybe we'll just make one node the authoritative attribute node, then
1350  * we can send this 'punch' to just the authoritative node and the nodes
1351  * that the punch will affect. */
1352 static int lov_punch(struct obd_export *exp, struct obdo *oa,
1353                      struct lov_stripe_md *lsm,
1354                      obd_off start, obd_off end, struct obd_trans_info *oti)
1355 {
1356         struct obdo tmp;
1357         struct lov_obd *lov;
1358         struct lov_oinfo *loi;
1359         int rc = 0, i;
1360         ENTRY;
1361
1362         if (lsm_bad_magic(lsm))
1363                 RETURN(-EINVAL);
1364
1365         if (!exp || !exp->exp_obd)
1366                 RETURN(-ENODEV);
1367
1368         lov = &exp->exp_obd->u.lov;
1369         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1370                 obd_off starti, endi;
1371                 int err;
1372
1373                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1374                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1375                         continue;
1376                 }
1377
1378                 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1379                         continue;
1380
1381                 /* create data objects with "parent" OA */
1382                 memcpy(&tmp, oa, sizeof(tmp));
1383                 tmp.o_id = loi->loi_id;
1384
1385                 err = obd_punch(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp, NULL,
1386                                 starti, endi, NULL);
1387                 if (err) {
1388                         if (lov->tgts[loi->loi_ost_idx].active) {
1389                                 CERROR("error: punch objid "LPX64" subobj "LPX64
1390                                        " on OST idx %d: rc = %d\n", oa->o_id,
1391                                        loi->loi_id, loi->loi_ost_idx, err);
1392                         }
1393                         if (!rc)
1394                                 rc = err;
1395                 } else {
1396                         loi->loi_kms = loi->loi_rss = starti;
1397                 }
1398         }
1399         RETURN(rc);
1400 }
1401
1402 static int lov_sync(struct obd_export *exp, struct obdo *oa,
1403                     struct lov_stripe_md *lsm, obd_off start, obd_off end)
1404 {
1405         struct obdo *tmp;
1406         struct lov_obd *lov;
1407         struct lov_oinfo *loi;
1408         int rc = 0, i;
1409         ENTRY;
1410
1411         if (lsm_bad_magic(lsm))
1412                 RETURN(-EINVAL);
1413
1414         if (!exp->exp_obd)
1415                 RETURN(-ENODEV);
1416
1417         tmp = obdo_alloc();
1418         if (!tmp)
1419                 RETURN(-ENOMEM);
1420
1421         lov = &exp->exp_obd->u.lov;
1422         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1423                 obd_off starti, endi;
1424                 int err;
1425
1426                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1427                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1428                         continue;
1429                 }
1430
1431                 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1432                         continue;
1433
1434                 memcpy(tmp, oa, sizeof(*tmp));
1435                 tmp->o_id = loi->loi_id;
1436
1437                 err = obd_sync(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp, NULL,
1438                                starti, endi);
1439                 if (err) {
1440                         if (lov->tgts[loi->loi_ost_idx].active) {
1441                                 CERROR("error: fsync objid "LPX64" subobj "LPX64
1442                                        " on OST idx %d: rc = %d\n", oa->o_id,
1443                                        loi->loi_id, loi->loi_ost_idx, err);
1444                         }
1445                         if (!rc)
1446                                 rc = err;
1447                 }
1448         }
1449
1450         obdo_free(tmp);
1451         RETURN(rc);
1452 }
1453
1454 static int lov_brw_check(struct lov_obd *lov, struct obdo *oa,
1455                          struct lov_stripe_md *lsm,
1456                          obd_count oa_bufs, struct brw_page *pga)
1457 {
1458         int i, rc = 0;
1459
1460         /* The caller just wants to know if there's a chance that this
1461          * I/O can succeed */
1462         for (i = 0; i < oa_bufs; i++) {
1463                 int stripe = lov_stripe_number(lsm, pga[i].off);
1464                 int ost = lsm->lsm_oinfo[stripe].loi_ost_idx;
1465                 struct ldlm_extent ext, subext;
1466                 ext.start = pga[i].off;
1467                 ext.end = pga[i].off + pga[i].count;
1468
1469                 if (!lov_stripe_intersects(lsm, i, ext.start, ext.end,
1470                                            &subext.start, &subext.end))
1471                         continue;
1472
1473                 if (lov->tgts[ost].active == 0) {
1474                         CDEBUG(D_HA, "lov idx %d inactive\n", ost);
1475                         return -EIO;
1476                 }
1477                 rc = obd_brw(OBD_BRW_CHECK, lov->tgts[stripe].ltd_exp, oa,
1478                              NULL, 1, &pga[i], NULL);
1479                 if (rc)
1480                         break;
1481         }
1482         return rc;
1483 }
1484
1485 static int lov_brw(int cmd, struct obd_export *exp, struct obdo *src_oa,
1486                    struct lov_stripe_md *lsm, obd_count oa_bufs,
1487                    struct brw_page *pga, struct obd_trans_info *oti)
1488 {
1489         struct {
1490                 int bufct;
1491                 int index;
1492                 int subcount;
1493                 struct lov_stripe_md lsm;
1494                 int ost_idx;
1495         } *stripeinfo, *si, *si_last;
1496         struct obdo *ret_oa = NULL, *tmp_oa = NULL;
1497         struct lov_obd *lov;
1498         struct brw_page *ioarr;
1499         struct lov_oinfo *loi;
1500         int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count, set = 0;
1501         ENTRY;
1502
1503         if (lsm_bad_magic(lsm))
1504                 RETURN(-EINVAL);
1505
1506         lov = &exp->exp_obd->u.lov;
1507
1508         if (cmd == OBD_BRW_CHECK) {
1509                 rc = lov_brw_check(lov, src_oa, lsm, oa_bufs, pga);
1510                 RETURN(rc);
1511         }
1512
1513         OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
1514         if (!stripeinfo)
1515                 RETURN(-ENOMEM);
1516
1517         OBD_ALLOC(where, sizeof(*where) * oa_bufs);
1518         if (!where)
1519                 GOTO(out_sinfo, rc = -ENOMEM);
1520
1521         OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
1522         if (!ioarr)
1523                 GOTO(out_where, rc = -ENOMEM);
1524
1525         if (src_oa) {
1526                 ret_oa = obdo_alloc();
1527                 if (!ret_oa)
1528                         GOTO(out_ioarr, rc = -ENOMEM);
1529
1530                 tmp_oa = obdo_alloc();
1531                 if (!tmp_oa)
1532                         GOTO(out_oa, rc = -ENOMEM);
1533         }
1534
1535         for (i = 0; i < oa_bufs; i++) {
1536                 where[i] = lov_stripe_number(lsm, pga[i].off);
1537                 stripeinfo[where[i]].bufct++;
1538         }
1539
1540         for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
1541              i < stripe_count; i++, loi++, si_last = si, si++) {
1542                 if (i > 0)
1543                         si->index = si_last->index + si_last->bufct;
1544                 si->lsm.lsm_object_id = loi->loi_id;
1545                 si->ost_idx = loi->loi_ost_idx;
1546         }
1547
1548         for (i = 0; i < oa_bufs; i++) {
1549                 int which = where[i];
1550                 int shift;
1551
1552                 shift = stripeinfo[which].index + stripeinfo[which].subcount;
1553                 LASSERT(shift < oa_bufs);
1554                 ioarr[shift] = pga[i];
1555                 lov_stripe_offset(lsm, pga[i].off, which, &ioarr[shift].off);
1556                 stripeinfo[which].subcount++;
1557         }
1558
1559         for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
1560                 int shift = si->index;
1561
1562                 if (lov->tgts[si->ost_idx].active == 0) {
1563                         CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
1564                         GOTO(out_oa, rc = -EIO);
1565                 }
1566
1567                 if (si->bufct) {
1568                         LASSERT(shift < oa_bufs);
1569                         if (src_oa)
1570                                 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1571
1572                         tmp_oa->o_id = si->lsm.lsm_object_id;
1573                         rc = obd_brw(cmd, lov->tgts[si->ost_idx].ltd_exp,
1574                                      tmp_oa, &si->lsm, si->bufct,
1575                                      &ioarr[shift], oti);
1576                         if (rc)
1577                                 GOTO(out_ioarr, rc);
1578
1579                         lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
1580                                         i, &set);
1581                 }
1582         }
1583
1584         ret_oa->o_id = src_oa->o_id;
1585         memcpy(src_oa, ret_oa, sizeof(*src_oa));
1586
1587         GOTO(out_oa, rc);
1588  out_oa:
1589         if (tmp_oa)
1590                 obdo_free(tmp_oa);
1591         if (ret_oa)
1592                 obdo_free(ret_oa);
1593  out_ioarr:
1594         OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1595  out_where:
1596         OBD_FREE(where, sizeof(*where) * oa_bufs);
1597  out_sinfo:
1598         OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
1599         return rc;
1600 }
1601
1602 static int lov_brw_interpret(struct ptlrpc_request_set *reqset, void *data,
1603                              int rc)
1604 {
1605         struct lov_brw_async_args *aa = data;
1606         struct lov_stripe_md *lsm = aa->aa_lsm;
1607         obd_count             oa_bufs = aa->aa_oa_bufs;
1608         struct obdo          *oa = aa->aa_oa;
1609         struct obdo          *obdos = aa->aa_obdos;
1610         struct brw_page      *ioarr = aa->aa_ioarr;
1611         struct lov_oinfo     *loi;
1612         int i, set = 0;
1613         ENTRY;
1614
1615         if (rc == 0) {
1616                 /* NB all stripe requests succeeded to get here */
1617
1618                 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1619                      i++, loi++) {
1620                         if (obdos[i].o_valid == 0)      /* inactive stripe */
1621                                 continue;
1622
1623                         lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1624                                         i, &set);
1625                 }
1626
1627                 if (!set) {
1628                         CERROR("No stripes had valid attrs\n");
1629                         rc = -EIO;
1630                 }
1631         }
1632         oa->o_id = lsm->lsm_object_id;
1633
1634         OBD_FREE(obdos, lsm->lsm_stripe_count * sizeof(*obdos));
1635         OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1636         RETURN(rc);
1637 }
1638
1639 static int lov_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1640                          struct lov_stripe_md *lsm, obd_count oa_bufs,
1641                          struct brw_page *pga, struct ptlrpc_request_set *set,
1642                          struct obd_trans_info *oti)
1643 {
1644         struct {
1645                 int bufct;
1646                 int index;
1647                 int subcount;
1648                 struct lov_stripe_md lsm;
1649                 int ost_idx;
1650         } *stripeinfo, *si, *si_last;
1651         struct lov_obd *lov;
1652         struct brw_page *ioarr;
1653         struct obdo *obdos = NULL;
1654         struct lov_oinfo *loi;
1655         struct lov_brw_async_args *aa;
1656         int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count;
1657         ENTRY;
1658
1659         if (lsm_bad_magic(lsm))
1660                 RETURN(-EINVAL);
1661
1662         lov = &exp->exp_obd->u.lov;
1663
1664         if (cmd == OBD_BRW_CHECK) {
1665                 rc = lov_brw_check(lov, oa, lsm, oa_bufs, pga);
1666                 RETURN(rc);
1667         }
1668
1669         OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
1670         if (!stripeinfo)
1671                 RETURN(-ENOMEM);
1672
1673         OBD_ALLOC(where, sizeof(*where) * oa_bufs);
1674         if (!where)
1675                 GOTO(out_sinfo, rc = -ENOMEM);
1676
1677         if (oa) {
1678                 OBD_ALLOC(obdos, sizeof(*obdos) * stripe_count);
1679                 if (!obdos)
1680                         GOTO(out_where, rc = -ENOMEM);
1681         }
1682
1683         OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
1684         if (!ioarr)
1685                 GOTO(out_obdos, rc = -ENOMEM);
1686
1687         for (i = 0; i < oa_bufs; i++) {
1688                 where[i] = lov_stripe_number(lsm, pga[i].off);
1689                 stripeinfo[where[i]].bufct++;
1690         }
1691
1692         for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
1693              i < stripe_count; i++, loi++, si_last = si, si++) {
1694                 if (i > 0)
1695                         si->index = si_last->index + si_last->bufct;
1696                 si->lsm.lsm_object_id = loi->loi_id;
1697                 si->ost_idx = loi->loi_ost_idx;
1698
1699                 if (oa) {
1700                         memcpy(&obdos[i], oa, sizeof(*obdos));
1701                         obdos[i].o_id = si->lsm.lsm_object_id;
1702                 }
1703         }
1704
1705         for (i = 0; i < oa_bufs; i++) {
1706                 int which = where[i];
1707                 int shift;
1708
1709                 shift = stripeinfo[which].index + stripeinfo[which].subcount;
1710                 LASSERT(shift < oa_bufs);
1711                 ioarr[shift] = pga[i];
1712                 lov_stripe_offset(lsm, pga[i].off, which, &ioarr[shift].off);
1713                 stripeinfo[which].subcount++;
1714         }
1715
1716         for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
1717                 int shift = si->index;
1718
1719                 if (si->bufct == 0)
1720                         continue;
1721
1722                 if (lov->tgts[si->ost_idx].active == 0) {
1723                         CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
1724                         GOTO(out_ioarr, rc = -EIO);
1725                 }
1726
1727                 LASSERT(shift < oa_bufs);
1728
1729                 rc = obd_brw_async(cmd, lov->tgts[si->ost_idx].ltd_exp,
1730                                    &obdos[i], &si->lsm, si->bufct,
1731                                    &ioarr[shift], set, oti);
1732                 if (rc)
1733                         GOTO(out_ioarr, rc);
1734         }
1735         LASSERT(rc == 0);
1736         LASSERT(set->set_interpret == NULL);
1737         set->set_interpret = (set_interpreter_func)lov_brw_interpret;
1738         LASSERT(sizeof(set->set_args) >= sizeof(struct lov_brw_async_args));
1739         aa = (struct lov_brw_async_args *)&set->set_args;
1740         aa->aa_lsm = lsm;
1741         aa->aa_obdos = obdos;
1742         aa->aa_oa = oa;
1743         aa->aa_ioarr = ioarr;
1744         aa->aa_oa_bufs = oa_bufs;
1745
1746         /* Don't free ioarr or obdos - that's done in lov_brw_interpret */
1747         GOTO(out_where, rc);
1748
1749  out_ioarr:
1750         OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1751  out_obdos:
1752         OBD_FREE(obdos, stripe_count * sizeof(*obdos));
1753  out_where:
1754         OBD_FREE(where, sizeof(*where) * oa_bufs);
1755  out_sinfo:
1756         OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
1757         return rc;
1758 }
1759
1760 struct lov_async_page *lap_from_cookie(void *cookie)
1761 {
1762         struct lov_async_page *lap = cookie;
1763         if (lap->lap_magic != LAP_MAGIC)
1764                 return ERR_PTR(-EINVAL);
1765         return lap;
1766 };
1767
1768 static int lov_ap_make_ready(void *data, int cmd)
1769 {
1770         struct lov_async_page *lap = lap_from_cookie(data);
1771         /* XXX should these assert? */
1772         if (IS_ERR(lap))
1773                 return -EINVAL;
1774
1775         return lap->lap_caller_ops->ap_make_ready(lap->lap_caller_data, cmd);
1776 }
1777 static int lov_ap_refresh_count(void *data, int cmd)
1778 {
1779         struct lov_async_page *lap = lap_from_cookie(data);
1780         if (IS_ERR(lap))
1781                 return -EINVAL;
1782
1783         return lap->lap_caller_ops->ap_refresh_count(lap->lap_caller_data,
1784                                                      cmd);
1785 }
1786 static void lov_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
1787 {
1788         struct lov_async_page *lap = lap_from_cookie(data);
1789         /* XXX should these assert? */
1790         if (IS_ERR(lap))
1791                 return;
1792
1793         lap->lap_caller_ops->ap_fill_obdo(lap->lap_caller_data, cmd, oa);
1794         /* XXX woah, shouldn't we be altering more here?  size? */
1795         oa->o_id = lap->lap_loi_id;
1796 }
1797
1798 static void lov_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
1799 {
1800         struct lov_async_page *lap = lap_from_cookie(data);
1801         if (IS_ERR(lap))
1802                 return;
1803
1804         /* in a raid1 regime this would down a count of many ios
1805          * in flight, onl calling the caller_ops completion when all
1806          * the raid1 ios are complete */
1807         lap->lap_caller_ops->ap_completion(lap->lap_caller_data, cmd, oa, rc);
1808 }
1809
1810 static struct obd_async_page_ops lov_async_page_ops = {
1811         .ap_make_ready =        lov_ap_make_ready,
1812         .ap_refresh_count =     lov_ap_refresh_count,
1813         .ap_fill_obdo =         lov_ap_fill_obdo,
1814         .ap_completion =        lov_ap_completion,
1815 };
1816
1817 int lov_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1818                            struct lov_oinfo *loi, struct page *page,
1819                            obd_off offset, struct obd_async_page_ops *ops,
1820                            void *data, void **res)
1821 {
1822         struct lov_obd *lov = &exp->exp_obd->u.lov;
1823         struct lov_async_page *lap;
1824         int rc;
1825         ENTRY;
1826
1827         if (lsm_bad_magic(lsm))
1828                 RETURN(-EINVAL);
1829         LASSERT(loi == NULL);
1830
1831         OBD_ALLOC(lap, sizeof(*lap));
1832         if (lap == NULL)
1833                 RETURN(-ENOMEM);
1834
1835         lap->lap_magic = LAP_MAGIC;
1836         lap->lap_caller_ops = ops;
1837         lap->lap_caller_data = data;
1838
1839         /* for now only raid 0 which passes through */
1840         lap->lap_stripe = lov_stripe_number(lsm, offset);
1841         lov_stripe_offset(lsm, offset, lap->lap_stripe, &lap->lap_sub_offset);
1842         loi = &lsm->lsm_oinfo[lap->lap_stripe];
1843
1844         /* so the callback doesn't need the lsm */
1845         lap->lap_loi_id = loi->loi_id;
1846
1847         rc = obd_prep_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
1848                                  lsm, loi, page, lap->lap_sub_offset,
1849                                  &lov_async_page_ops, lap,
1850                                  &lap->lap_sub_cookie);
1851         if (rc) {
1852                 OBD_FREE(lap, sizeof(*lap));
1853                 RETURN(rc);
1854         }
1855         CDEBUG(D_CACHE, "lap %p page %p cookie %p off "LPU64"\n", lap, page,
1856                lap->lap_sub_cookie, offset);
1857         *res = lap;
1858         RETURN(0);
1859 }
1860
1861 static int lov_queue_async_io(struct obd_export *exp,
1862                               struct lov_stripe_md *lsm,
1863                               struct lov_oinfo *loi, void *cookie,
1864                               int cmd, obd_off off, int count,
1865                               obd_flag brw_flags, obd_flag async_flags)
1866 {
1867         struct lov_obd *lov = &exp->exp_obd->u.lov;
1868         struct lov_async_page *lap;
1869         int rc;
1870
1871         LASSERT(loi == NULL);
1872
1873         if (lsm_bad_magic(lsm))
1874                 RETURN(-EINVAL);
1875
1876         lap = lap_from_cookie(cookie);
1877         if (IS_ERR(lap))
1878                 RETURN(PTR_ERR(lap));
1879
1880         loi = &lsm->lsm_oinfo[lap->lap_stripe];
1881         rc = obd_queue_async_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm,
1882                                 loi, lap->lap_sub_cookie, cmd, off, count,
1883                                 brw_flags, async_flags);
1884         RETURN(rc);
1885 }
1886
1887 static int lov_set_async_flags(struct obd_export *exp,
1888                                struct lov_stripe_md *lsm,
1889                                struct lov_oinfo *loi, void *cookie,
1890                                obd_flag async_flags)
1891 {
1892         struct lov_obd *lov = &exp->exp_obd->u.lov;
1893         struct lov_async_page *lap;
1894         int rc;
1895
1896         LASSERT(loi == NULL);
1897
1898         if (lsm_bad_magic(lsm))
1899                 RETURN(-EINVAL);
1900
1901         lap = lap_from_cookie(cookie);
1902         if (IS_ERR(lap))
1903                 RETURN(PTR_ERR(lap));
1904
1905         loi = &lsm->lsm_oinfo[lap->lap_stripe];
1906         rc = obd_set_async_flags(lov->tgts[loi->loi_ost_idx].ltd_exp,
1907                                  lsm, loi, lap->lap_sub_cookie, async_flags);
1908         RETURN(rc);
1909 }
1910
1911 static int lov_queue_group_io(struct obd_export *exp,
1912                               struct lov_stripe_md *lsm,
1913                               struct lov_oinfo *loi,
1914                               struct obd_io_group *oig, void *cookie,
1915                               int cmd, obd_off off, int count,
1916                               obd_flag brw_flags, obd_flag async_flags)
1917 {
1918         struct lov_obd *lov = &exp->exp_obd->u.lov;
1919         struct lov_async_page *lap;
1920         int rc;
1921
1922         LASSERT(loi == NULL);
1923
1924         if (lsm_bad_magic(lsm))
1925                 RETURN(-EINVAL);
1926
1927         lap = lap_from_cookie(cookie);
1928         if (IS_ERR(lap))
1929                 RETURN(PTR_ERR(lap));
1930
1931         loi = &lsm->lsm_oinfo[lap->lap_stripe];
1932         rc = obd_queue_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm, loi,
1933                                 oig, lap->lap_sub_cookie, cmd, off, count,
1934                                 brw_flags, async_flags);
1935         RETURN(rc);
1936 }
1937
1938 /* this isn't exactly optimal.  we may have queued sync io in oscs on
1939  * all stripes, but we don't record that fact at queue time.  so we
1940  * trigger sync io on all stripes. */
1941 static int lov_trigger_group_io(struct obd_export *exp,
1942                                 struct lov_stripe_md *lsm,
1943                                 struct lov_oinfo *loi,
1944                                 struct obd_io_group *oig)
1945 {
1946         struct lov_obd *lov = &exp->exp_obd->u.lov;
1947         int rc = 0, i, err;
1948
1949         LASSERT(loi == NULL);
1950
1951         if (lsm_bad_magic(lsm))
1952                 RETURN(-EINVAL);
1953
1954         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1955              i++, loi++) {
1956                 err = obd_trigger_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp,
1957                                            lsm, loi, oig);
1958                 if (rc == 0 && err != 0)
1959                         rc = err;
1960         };
1961         RETURN(rc);
1962 }
1963
1964 static int lov_teardown_async_page(struct obd_export *exp,
1965                                    struct lov_stripe_md *lsm,
1966                                    struct lov_oinfo *loi, void *cookie)
1967 {
1968         struct lov_obd *lov = &exp->exp_obd->u.lov;
1969         struct lov_async_page *lap;
1970         int rc;
1971
1972         LASSERT(loi == NULL);
1973
1974         if (lsm_bad_magic(lsm))
1975                 RETURN(-EINVAL);
1976
1977         lap = lap_from_cookie(cookie);
1978         if (IS_ERR(lap))
1979                 RETURN(PTR_ERR(lap));
1980
1981         loi = &lsm->lsm_oinfo[lap->lap_stripe];
1982         rc = obd_teardown_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
1983                                      lsm, loi, lap->lap_sub_cookie);
1984         if (rc) {
1985                 CERROR("unable to teardown sub cookie %p: %d\n",
1986                        lap->lap_sub_cookie, rc);
1987                 RETURN(rc);
1988         }
1989         OBD_FREE(lap, sizeof(*lap));
1990         RETURN(rc);
1991 }
1992
1993 static int lov_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
1994                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
1995                        int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
1996                        void *data,__u32 lvb_len, void *lvb_swabber,
1997                        struct lustre_handle *lockh)
1998 {
1999         struct lov_lock_handles *lov_lockh = NULL;
2000         struct lustre_handle *lov_lockhp;
2001         struct lov_obd *lov;
2002         struct lov_oinfo *loi;
2003         char submd_buf[sizeof(struct lov_stripe_md) + sizeof(struct lov_oinfo)];
2004         struct lov_stripe_md *submd = (void *)submd_buf;
2005         ldlm_error_t rc;
2006         int i, save_flags = *flags;
2007         ENTRY;
2008
2009         if (lsm_bad_magic(lsm))
2010                 RETURN(-EINVAL);
2011
2012         /* we should never be asked to replay a lock this way. */
2013         LASSERT((*flags & LDLM_FL_REPLAY) == 0);
2014
2015         if (!exp || !exp->exp_obd)
2016                 RETURN(-ENODEV);
2017
2018         if (lsm->lsm_stripe_count > 1) {
2019                 lov_lockh = lov_llh_new(lsm);
2020                 if (lov_lockh == NULL)
2021                         RETURN(-ENOMEM);
2022
2023                 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2024                 lov_lockhp = lov_lockh->llh_handles;
2025         } else {
2026                 lov_lockhp = lockh;
2027         }
2028
2029         lov = &exp->exp_obd->u.lov;
2030         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2031              i++, loi++, lov_lockhp++) {
2032                 ldlm_policy_data_t sub_ext;
2033
2034                 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2035                                            policy->l_extent.end,
2036                                            &sub_ext.l_extent.start,
2037                                            &sub_ext.l_extent.end))
2038                         continue;
2039
2040                 sub_ext.l_extent.gid = policy->l_extent.gid;
2041
2042                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2043                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2044                         continue;
2045                 }
2046
2047                 /* XXX LOV STACKING: submd should be from the subobj */
2048                 submd->lsm_object_id = loi->loi_id;
2049                 submd->lsm_stripe_count = 0;
2050                 submd->lsm_oinfo->loi_kms_valid = loi->loi_kms_valid;
2051                 submd->lsm_oinfo->loi_rss = loi->loi_rss;
2052                 submd->lsm_oinfo->loi_kms = loi->loi_kms;
2053                 submd->lsm_oinfo->loi_blocks = loi->loi_blocks;
2054                 loi->loi_mtime = submd->lsm_oinfo->loi_mtime;
2055                 /* XXX submd is not fully initialized here */
2056                 *flags = save_flags;
2057                 rc = obd_enqueue(lov->tgts[loi->loi_ost_idx].ltd_exp, submd,
2058                                  type, &sub_ext, mode, flags, bl_cb, cp_cb,
2059                                  gl_cb, data, lvb_len, lvb_swabber, lov_lockhp);
2060
2061                 /* XXX FIXME: This unpleasantness doesn't belong here at *all*.
2062                  * It belongs in the OSC, except that the OSC doesn't have
2063                  * access to the real LOI -- it gets a copy, that we created
2064                  * above, and that copy can be arbitrarily out of date.
2065                  *
2066                  * The LOV API is due for a serious rewriting anyways, and this
2067                  * can be addressed then. */
2068                 if (rc == ELDLM_OK) {
2069                         struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2070                         __u64 tmp = submd->lsm_oinfo->loi_rss;
2071
2072                         LASSERT(lock != NULL);
2073                         loi->loi_rss = tmp;
2074                         loi->loi_blocks = submd->lsm_oinfo->loi_blocks;
2075                         /* Extend KMS up to the end of this lock and no further
2076                          * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2077                         if (tmp > lock->l_policy_data.l_extent.end)
2078                                 tmp = lock->l_policy_data.l_extent.end + 1;
2079                         if (tmp >= loi->loi_kms) {
2080                                 CDEBUG(D_INODE, "lock acquired, setting rss="
2081                                        LPU64", kms="LPU64"\n", loi->loi_rss,
2082                                        tmp);
2083                                 loi->loi_kms = tmp;
2084                                 loi->loi_kms_valid = 1;
2085                         } else {
2086                                 CDEBUG(D_INODE, "lock acquired, setting rss="
2087                                        LPU64"; leaving kms="LPU64", end="LPU64
2088                                        "\n", loi->loi_rss, loi->loi_kms,
2089                                        lock->l_policy_data.l_extent.end);
2090                         }
2091                         ldlm_lock_allow_match(lock);
2092                         LDLM_LOCK_PUT(lock);
2093                 } else if (rc == ELDLM_LOCK_ABORTED &&
2094                            save_flags & LDLM_FL_HAS_INTENT) {
2095                         memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2096                         loi->loi_rss = submd->lsm_oinfo->loi_rss;
2097                         loi->loi_blocks = submd->lsm_oinfo->loi_blocks;
2098                         CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2099                                " kms="LPU64"\n", loi->loi_rss, loi->loi_kms);
2100                 } else {
2101                         memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2102                         if (lov->tgts[loi->loi_ost_idx].active) {
2103                                 CERROR("error: enqueue objid "LPX64" subobj "
2104                                        LPX64" on OST idx %d: rc = %d\n",
2105                                        lsm->lsm_object_id, loi->loi_id,
2106                                        loi->loi_ost_idx, rc);
2107                                 GOTO(out_locks, rc);
2108                         }
2109                 }
2110         }
2111         if (lsm->lsm_stripe_count > 1)
2112                 lov_llh_put(lov_lockh);
2113         RETURN(ELDLM_OK);
2114
2115  out_locks:
2116         while (loi--, lov_lockhp--, i-- > 0) {
2117                 struct lov_stripe_md submd;
2118                 int err;
2119
2120                 if (lov_lockhp->cookie == 0)
2121                         continue;
2122
2123                 /* XXX LOV STACKING: submd should be from the subobj */
2124                 submd.lsm_object_id = loi->loi_id;
2125                 submd.lsm_stripe_count = 0;
2126                 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2127                                  mode, lov_lockhp);
2128                 if (err && lov->tgts[loi->loi_ost_idx].active) {
2129                         CERROR("error: cancelling objid "LPX64" on OST "
2130                                "idx %d after enqueue error: rc = %d\n",
2131                                loi->loi_id, loi->loi_ost_idx, err);
2132                 }
2133         }
2134
2135         if (lsm->lsm_stripe_count > 1) {
2136                 lov_llh_destroy(lov_lockh);
2137                 lov_llh_put(lov_lockh);
2138         }
2139         return rc;
2140 }
2141
2142 static int lov_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2143                      __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2144                      int *flags, void *data, struct lustre_handle *lockh)
2145 {
2146         struct lov_lock_handles *lov_lockh = NULL;
2147         struct lustre_handle *lov_lockhp;
2148         struct lov_obd *lov;
2149         struct lov_oinfo *loi;
2150         struct lov_stripe_md submd;
2151         ldlm_error_t rc = 0;
2152         int i;
2153         ENTRY;
2154
2155         if (lsm_bad_magic(lsm))
2156                 RETURN(-EINVAL);
2157
2158         if (!exp || !exp->exp_obd)
2159                 RETURN(-ENODEV);
2160
2161         if (lsm->lsm_stripe_count > 1) {
2162                 lov_lockh = lov_llh_new(lsm);
2163                 if (lov_lockh == NULL)
2164                         RETURN(-ENOMEM);
2165
2166                 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2167                 lov_lockhp = lov_lockh->llh_handles;
2168         } else {
2169                 lov_lockhp = lockh;
2170         }
2171
2172         lov = &exp->exp_obd->u.lov;
2173         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2174              i++, loi++, lov_lockhp++) {
2175                 ldlm_policy_data_t sub_ext;
2176                 int lov_flags;
2177
2178                 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2179                                            policy->l_extent.end,
2180                                            &sub_ext.l_extent.start,
2181                                            &sub_ext.l_extent.end))
2182                         continue;
2183
2184                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2185                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2186                         rc = -EIO;
2187                         break;
2188                 }
2189
2190                 /* XXX LOV STACKING: submd should be from the subobj */
2191                 submd.lsm_object_id = loi->loi_id;
2192                 submd.lsm_stripe_count = 0;
2193                 lov_flags = *flags;
2194                 /* XXX submd is not fully initialized here */
2195                 rc = obd_match(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2196                                type, &sub_ext, mode, &lov_flags, data,
2197                                lov_lockhp);
2198                 if (rc != 1)
2199                         break;
2200         }
2201         if (rc == 1) {
2202                 if (lsm->lsm_stripe_count > 1) {
2203                         if (*flags & LDLM_FL_TEST_LOCK)
2204                                 lov_llh_destroy(lov_lockh);
2205                         lov_llh_put(lov_lockh);
2206                 }
2207                 RETURN(1);
2208         }
2209
2210         while (loi--, lov_lockhp--, i-- > 0) {
2211                 struct lov_stripe_md submd;
2212                 int err;
2213
2214                 if (lov_lockhp->cookie == 0)
2215                         continue;
2216
2217                 /* XXX LOV STACKING: submd should be from the subobj */
2218                 submd.lsm_object_id = loi->loi_id;
2219                 submd.lsm_stripe_count = 0;
2220                 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2221                                  mode, lov_lockhp);
2222                 if (err && lov->tgts[loi->loi_ost_idx].active) {
2223                         CERROR("error: cancelling objid "LPX64" on OST "
2224                                "idx %d after match failure: rc = %d\n",
2225                                loi->loi_id, loi->loi_ost_idx, err);
2226                 }
2227         }
2228
2229         if (lsm->lsm_stripe_count > 1) {
2230                 lov_llh_destroy(lov_lockh);
2231                 lov_llh_put(lov_lockh);
2232         }
2233         RETURN(rc);
2234 }
2235
2236 static int lov_change_cbdata(struct obd_export *exp,
2237                              struct lov_stripe_md *lsm, ldlm_iterator_t it,
2238                              void *data)
2239 {
2240         struct lov_obd *lov;
2241         struct lov_oinfo *loi;
2242         int rc = 0, i;
2243         ENTRY;
2244
2245         if (lsm_bad_magic(lsm))
2246                 RETURN(-EINVAL);
2247
2248         if (!exp || !exp->exp_obd)
2249                 RETURN(-ENODEV);
2250
2251         lov = &exp->exp_obd->u.lov;
2252         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2253                 struct lov_stripe_md submd;
2254                 if (lov->tgts[loi->loi_ost_idx].active == 0)
2255                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2256
2257                 submd.lsm_object_id = loi->loi_id;
2258                 submd.lsm_stripe_count = 0;
2259                 rc = obd_change_cbdata(lov->tgts[loi->loi_ost_idx].ltd_exp,
2260                                        &submd, it, data);
2261         }
2262         RETURN(rc);
2263 }
2264
2265 static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm,
2266                       __u32 mode, struct lustre_handle *lockh)
2267 {
2268         struct lov_lock_handles *lov_lockh = NULL;
2269         struct lustre_handle *lov_lockhp;
2270         struct lov_obd *lov;
2271         struct lov_oinfo *loi;
2272         int rc = 0, i;
2273         ENTRY;
2274
2275         if (lsm_bad_magic(lsm))
2276                 RETURN(-EINVAL);
2277
2278         if (!exp || !exp->exp_obd)
2279                 RETURN(-ENODEV);
2280
2281         LASSERT(lockh);
2282         if (lsm->lsm_stripe_count > 1) {
2283                 lov_lockh = lov_handle2llh(lockh);
2284                 if (!lov_lockh) {
2285                         CERROR("LOV: invalid lov lock handle %p\n", lockh);
2286                         RETURN(-EINVAL);
2287                 }
2288
2289                 lov_lockhp = lov_lockh->llh_handles;
2290         } else {
2291                 lov_lockhp = lockh;
2292         }
2293
2294         lov = &exp->exp_obd->u.lov;
2295         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2296              i++, loi++, lov_lockhp++) {
2297                 struct lov_stripe_md submd;
2298                 int err;
2299
2300                 if (lov_lockhp->cookie == 0) {
2301                         CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2302                                loi->loi_ost_idx, loi->loi_id);
2303                         continue;
2304                 }
2305
2306                 /* XXX LOV STACKING: submd should be from the subobj */
2307                 submd.lsm_object_id = loi->loi_id;
2308                 submd.lsm_stripe_count = 0;
2309                 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2310                                  mode, lov_lockhp);
2311                 if (err) {
2312                         if (lov->tgts[loi->loi_ost_idx].active) {
2313                                 CERROR("error: cancel objid "LPX64" subobj "
2314                                        LPX64" on OST idx %d: rc = %d\n",
2315                                        lsm->lsm_object_id,
2316                                        loi->loi_id, loi->loi_ost_idx, err);
2317                                 if (!rc)
2318                                         rc = err;
2319                         }
2320                 }
2321         }
2322
2323         if (lsm->lsm_stripe_count > 1)
2324                 lov_llh_destroy(lov_lockh);
2325         if (lov_lockh != NULL)
2326                 lov_llh_put(lov_lockh);
2327         RETURN(rc);
2328 }
2329
2330 static int lov_cancel_unused(struct obd_export *exp,
2331                              struct lov_stripe_md *lsm, int flags, void *opaque)
2332 {
2333         struct lov_obd *lov;
2334         struct lov_oinfo *loi;
2335         int rc = 0, i;
2336         ENTRY;
2337
2338         if (lsm_bad_magic(lsm))
2339                 RETURN(-EINVAL);
2340
2341         if (!exp || !exp->exp_obd)
2342                 RETURN(-ENODEV);
2343
2344         lov = &exp->exp_obd->u.lov;
2345         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2346                 struct lov_stripe_md submd;
2347                 int err;
2348
2349                 if (lov->tgts[loi->loi_ost_idx].active == 0)
2350                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2351
2352                 submd.lsm_object_id = loi->loi_id;
2353                 submd.lsm_stripe_count = 0;
2354                 err = obd_cancel_unused(lov->tgts[loi->loi_ost_idx].ltd_exp,
2355                                         &submd, flags, opaque);
2356                 if (err && lov->tgts[loi->loi_ost_idx].active) {
2357                         CERROR("error: cancel unused objid "LPX64" subobj "LPX64
2358                                " on OST idx %d: rc = %d\n", lsm->lsm_object_id,
2359                                loi->loi_id, loi->loi_ost_idx, err);
2360                         if (!rc)
2361                                 rc = err;
2362                 }
2363         }
2364         RETURN(rc);
2365 }
2366
2367 #define LOV_U64_MAX ((__u64)~0ULL)
2368 #define LOV_SUM_MAX(tot, add)                                           \
2369         do {                                                            \
2370                 if ((tot) + (add) < (tot))                              \
2371                         (tot) = LOV_U64_MAX;                            \
2372                 else                                                    \
2373                         (tot) += (add);                                 \
2374         } while(0)
2375
2376 static int lov_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2377                       unsigned long max_age)
2378 {
2379         struct lov_obd *lov = &obd->u.lov;
2380         struct obd_statfs lov_sfs;
2381         int set = 0;
2382         int rc = 0;
2383         int i;
2384         ENTRY;
2385
2386
2387         /* We only get block data from the OBD */
2388         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2389                 int err;
2390
2391                 if (!lov->tgts[i].active) {
2392                         CDEBUG(D_HA, "lov idx %d inactive\n", i);
2393                         continue;
2394                 }
2395
2396                 err = obd_statfs(class_exp2obd(lov->tgts[i].ltd_exp), &lov_sfs,
2397                                  max_age);
2398                 if (err) {
2399                         if (lov->tgts[i].active && !rc)
2400                                 rc = err;
2401                         continue;
2402                 }
2403
2404                 if (!set) {
2405                         memcpy(osfs, &lov_sfs, sizeof(lov_sfs));
2406                         set = 1;
2407                 } else {
2408                         osfs->os_bfree += lov_sfs.os_bfree;
2409                         osfs->os_bavail += lov_sfs.os_bavail;
2410                         osfs->os_blocks += lov_sfs.os_blocks;
2411                         /* XXX not sure about this one - depends on policy.
2412                          *   - could be minimum if we always stripe on all OBDs
2413                          *     (but that would be wrong for any other policy,
2414                          *     if one of the OBDs has no more objects left)
2415                          *   - could be sum if we stripe whole objects
2416                          *   - could be average, just to give a nice number
2417                          *
2418                          * To give a "reasonable" (if not wholly accurate)
2419                          * number, we divide the total number of free objects
2420                          * by expected stripe count (watch out for overflow).
2421                          */
2422                         LOV_SUM_MAX(osfs->os_files, lov_sfs.os_files);
2423                         LOV_SUM_MAX(osfs->os_ffree, lov_sfs.os_ffree);
2424                 }
2425         }
2426
2427         if (set) {
2428                 __u32 expected_stripes = lov->desc.ld_default_stripe_count ?
2429                                          lov->desc.ld_default_stripe_count :
2430                                          lov->desc.ld_active_tgt_count;
2431
2432                 if (osfs->os_files != LOV_U64_MAX)
2433                         do_div(osfs->os_files, expected_stripes);
2434                 if (osfs->os_ffree != LOV_U64_MAX)
2435                         do_div(osfs->os_ffree, expected_stripes);
2436         } else if (!rc)
2437                 rc = -EIO;
2438
2439         RETURN(rc);
2440 }
2441
2442 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2443                          void *karg, void *uarg)
2444 {
2445         struct obd_device *obddev = class_exp2obd(exp);
2446         struct lov_obd *lov = &obddev->u.lov;
2447         int i, count = lov->desc.ld_tgt_count;
2448         struct obd_uuid *uuidp;
2449         int rc;
2450
2451         ENTRY;
2452
2453         switch (cmd) {
2454         case OBD_IOC_LOV_GET_CONFIG: {
2455                 struct obd_ioctl_data *data = karg;
2456                 struct lov_tgt_desc *tgtdesc;
2457                 struct lov_desc *desc;
2458                 char *buf = NULL;
2459
2460                 buf = NULL;
2461                 len = 0;
2462                 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2463                         RETURN(-EINVAL);
2464
2465                 data = (struct obd_ioctl_data *)buf;
2466
2467                 if (sizeof(*desc) > data->ioc_inllen1) {
2468                         OBD_FREE(buf, len);
2469                         RETURN(-EINVAL);
2470                 }
2471
2472                 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
2473                         OBD_FREE(buf, len);
2474                         RETURN(-EINVAL);
2475                 }
2476
2477                 desc = (struct lov_desc *)data->ioc_inlbuf1;
2478                 memcpy(desc, &(lov->desc), sizeof(*desc));
2479
2480                 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
2481                 tgtdesc = lov->tgts;
2482                 for (i = 0; i < count; i++, uuidp++, tgtdesc++)
2483                         obd_str2uuid(uuidp, tgtdesc->uuid.uuid);
2484
2485                 rc = copy_to_user((void *)uarg, buf, len);
2486                 if (rc)
2487                         rc = -EFAULT;
2488                 obd_ioctl_freedata(buf, len);
2489                 break;
2490         }
2491         case LL_IOC_LOV_SETSTRIPE:
2492                 rc = lov_setstripe(exp, karg, uarg);
2493                 break;
2494         case LL_IOC_LOV_GETSTRIPE:
2495                 rc = lov_getstripe(exp, karg, uarg);
2496                 break;
2497         case LL_IOC_LOV_SETEA:
2498                 rc = lov_setea(exp, karg, uarg);
2499                 break;
2500         default: {
2501                 int set = 0;
2502                 if (count == 0)
2503                         RETURN(-ENOTTY);
2504                 rc = 0;
2505                 for (i = 0; i < count; i++) {
2506                         int err;
2507
2508                         err = obd_iocontrol(cmd, lov->tgts[i].ltd_exp,
2509                                             len, karg, uarg);
2510                         if (err) {
2511                                 if (lov->tgts[i].active) {
2512                                         CERROR("error: iocontrol OSC %s on OST"
2513                                                "idx %d: err = %d\n",
2514                                                lov->tgts[i].uuid.uuid, i, err);
2515                                         if (!rc)
2516                                                 rc = err;
2517                                 }
2518                         } else
2519                                 set = 1;
2520                 }
2521                 if (!set && !rc)
2522                         rc = -EIO;
2523         }
2524         }
2525
2526         RETURN(rc);
2527 }
2528
2529 static int lov_get_info(struct obd_export *exp, __u32 keylen,
2530                         void *key, __u32 *vallen, void *val)
2531 {
2532         struct obd_device *obddev = class_exp2obd(exp);
2533         struct lov_obd *lov = &obddev->u.lov;
2534         int i;
2535         ENTRY;
2536
2537         if (!vallen || !val)
2538                 RETURN(-EFAULT);
2539
2540         if (keylen > strlen("lock_to_stripe") &&
2541             strcmp(key, "lock_to_stripe") == 0) {
2542                 struct {
2543                         char name[16];
2544                         struct ldlm_lock *lock;
2545                         struct lov_stripe_md *lsm;
2546                 } *data = key;
2547                 struct lov_oinfo *loi;
2548                 __u32 *stripe = val;
2549
2550                 if (*vallen < sizeof(*stripe))
2551                         RETURN(-EFAULT);
2552                 *vallen = sizeof(*stripe);
2553
2554                 /* XXX This is another one of those bits that will need to
2555                  * change if we ever actually support nested LOVs.  It uses
2556                  * the lock's export to find out which stripe it is. */
2557                 for (i = 0, loi = data->lsm->lsm_oinfo;
2558                      i < data->lsm->lsm_stripe_count;
2559                      i++, loi++) {
2560                         if (lov->tgts[loi->loi_ost_idx].ltd_exp ==
2561                             data->lock->l_conn_export) {
2562                                 *stripe = i;
2563                                 RETURN(0);
2564                         }
2565                 }
2566                 RETURN(-ENXIO);
2567         } else if (keylen >= strlen("size_to_stripe") &&
2568                    strcmp(key, "size_to_stripe") == 0) {
2569                 struct {
2570                         int stripe_number;
2571                         __u64 size;
2572                         struct lov_stripe_md *lsm;
2573                 } *data = val;
2574
2575                 if (*vallen < sizeof(*data))
2576                         RETURN(-EFAULT);
2577
2578                 data->size = lov_size_to_stripe(data->lsm, data->size,
2579                                                 data->stripe_number);
2580                 RETURN(0);
2581         } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2582                 obd_id *ids = val;
2583                 int rc, size = sizeof(obd_id);
2584                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2585                         if (!lov->tgts[i].active)
2586                                 continue;
2587                         rc = obd_get_info(lov->tgts[i].ltd_exp, keylen, key,
2588                                           &size, &(ids[i]));
2589                         if (rc != 0)
2590                                 RETURN(rc);
2591                 }
2592                 RETURN(0);
2593         } else if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
2594                 struct lov_desc *desc_ret = val;
2595                 *desc_ret = lov->desc;
2596
2597                 RETURN(0);
2598         }
2599
2600         RETURN(-EINVAL);
2601 }
2602
2603 static int lov_set_info(struct obd_export *exp, obd_count keylen,
2604                         void *key, obd_count vallen, void *val)
2605 {
2606         struct obd_device *obddev = class_exp2obd(exp);
2607         struct lov_obd *lov = &obddev->u.lov;
2608         int i, rc = 0;
2609         ENTRY;
2610
2611 #define KEY_IS(str) \
2612         (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
2613
2614         if (KEY_IS("next_id")) {
2615                 if (vallen != lov->desc.ld_tgt_count)
2616                         RETURN(-EINVAL);
2617                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2618                         int er;
2619
2620                         /* initialize all OSCs, even inactive ones */
2621
2622                         er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key,
2623                                           sizeof(obd_id), ((obd_id*)val) + i);
2624                         if (!rc)
2625                                 rc = er;
2626                 }
2627                 RETURN(rc);
2628         }
2629
2630         if (KEY_IS("growth_count")) {
2631                 if (vallen != sizeof(int))
2632                         RETURN(-EINVAL);
2633         } else if (KEY_IS("mds_conn") || KEY_IS("unlinked")) {
2634                 if (vallen != 0)
2635                         RETURN(-EINVAL);
2636         } else {
2637                 RETURN(-EINVAL);
2638         }
2639
2640         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2641                 int er;
2642
2643                 if (val && !obd_uuid_equals(val, &lov->tgts[i].uuid))
2644                         continue;
2645
2646                 if (!val && !lov->tgts[i].active)
2647                         continue;
2648
2649                 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key, vallen,
2650                                    val);
2651                 if (!rc)
2652                         rc = er;
2653         }
2654         RETURN(rc);
2655 #undef KEY_IS
2656
2657 }
2658
2659 /* Merge rss if kms == 0
2660  *
2661  * Even when merging RSS, we will take the KMS value if it's larger.
2662  * This prevents getattr from stomping on dirty cached pages which
2663  * extend the file size. */
2664 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms)
2665 {
2666         struct lov_oinfo *loi;
2667         __u64 size = 0;
2668         int i;
2669
2670         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2671              i++, loi++) {
2672                 obd_size lov_size, tmpsize;
2673
2674                 tmpsize = loi->loi_kms;
2675                 if (kms == 0 && loi->loi_rss > tmpsize)
2676                         tmpsize = loi->loi_rss;
2677
2678                 lov_size = lov_stripe_size(lsm, tmpsize, i);
2679                 if (lov_size > size)
2680                         size = lov_size;
2681         }
2682
2683         return size;
2684 }
2685 EXPORT_SYMBOL(lov_merge_size);
2686
2687 /* Merge blocks */
2688 __u64 lov_merge_blocks(struct lov_stripe_md *lsm)
2689 {
2690         struct lov_oinfo *loi;
2691         __u64 blocks = 0;
2692         int i;
2693
2694         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2695              i++, loi++) {
2696                 blocks += loi->loi_blocks;
2697         }
2698         return blocks;
2699 }
2700 EXPORT_SYMBOL(lov_merge_blocks);
2701
2702 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time)
2703 {
2704         struct lov_oinfo *loi;
2705         int i;
2706
2707         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2708              i++, loi++) {
2709                 if (loi->loi_mtime > current_time)
2710                         current_time = loi->loi_mtime;
2711         }
2712         return current_time;
2713 }
2714 EXPORT_SYMBOL(lov_merge_mtime);
2715
2716 #if 0
2717 struct lov_multi_wait {
2718         struct ldlm_lock *lock;
2719         wait_queue_t      wait;
2720         int               completed;
2721         int               generation;
2722 };
2723
2724 int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm,
2725                       struct lustre_handle *lockh)
2726 {
2727         struct lov_lock_handles *lov_lockh = NULL;
2728         struct lustre_handle *lov_lockhp;
2729         struct lov_obd *lov;
2730         struct lov_oinfo *loi;
2731         struct lov_multi_wait *queues;
2732         int rc = 0, i;
2733         ENTRY;
2734
2735         if (lsm_bad_magic(lsm))
2736                 RETURN(-EINVAL);
2737
2738         if (!exp || !exp->exp_obd)
2739                 RETURN(-ENODEV);
2740
2741         LASSERT(lockh != NULL);
2742         if (lsm->lsm_stripe_count > 1) {
2743                 lov_lockh = lov_handle2llh(lockh);
2744                 if (lov_lockh == NULL) {
2745                         CERROR("LOV: invalid lov lock handle %p\n", lockh);
2746                         RETURN(-EINVAL);
2747                 }
2748
2749                 lov_lockhp = lov_lockh->llh_handles;
2750         } else {
2751                 lov_lockhp = lockh;
2752         }
2753
2754         OBD_ALLOC(queues, lsm->lsm_stripe_count * sizeof(*queues));
2755         if (queues == NULL)
2756                 GOTO(out, rc = -ENOMEM);
2757
2758         lov = &exp->exp_obd->u.lov;
2759         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2760              i++, loi++, lov_lockhp++) {
2761                 struct ldlm_lock *lock;
2762                 struct obd_device *obd;
2763                 unsigned long irqflags;
2764
2765                 lock = ldlm_handle2lock(lov_lockhp);
2766                 if (lock == NULL) {
2767                         CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2768                                loi->loi_ost_idx, loi->loi_id);
2769                         queues[i].completed = 1;
2770                         continue;
2771                 }
2772
2773                 queues[i].lock = lock;
2774                 init_waitqueue_entry(&(queues[i].wait), current);
2775                 add_wait_queue(lock->l_waitq, &(queues[i].wait));
2776
2777                 obd = class_exp2obd(lock->l_conn_export);
2778                 if (obd != NULL)
2779                         imp = obd->u.cli.cl_import;
2780                 if (imp != NULL) {
2781                         spin_lock_irqsave(&imp->imp_lock, irqflags);
2782                         queues[i].generation = imp->imp_generation;
2783                         spin_unlock_irqrestore(&imp->imp_lock, irqflags);
2784                 }
2785         }
2786
2787         lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
2788                                interrupted_completion_wait, &lwd);
2789         rc = l_wait_event_added(check_multi_complete(queues, lsm), &lwi);
2790
2791         for (i = 0; i < lsm->lsm_stripe_count; i++)
2792                 remove_wait_queue(lock->l_waitq, &(queues[i].wait));
2793
2794         if (rc == -EINTR || rc == -ETIMEDOUT) {
2795
2796
2797         }
2798
2799  out:
2800         if (lov_lockh != NULL)
2801                 lov_llh_put(lov_lockh);
2802         RETURN(rc);
2803 }
2804 #endif
2805
2806 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
2807                       obd_off size)
2808 {
2809         struct lov_oinfo *loi;
2810         int stripe = 0;
2811         __u64 kms;
2812         ENTRY;
2813
2814         if (size > 0)
2815                 stripe = lov_stripe_number(lsm, size - 1);
2816         kms = lov_size_to_stripe(lsm, size, stripe);
2817         loi = &(lsm->lsm_oinfo[stripe]);
2818
2819         CDEBUG(D_INODE, "stripe %d KMS %sincreasing "LPU64"->"LPU64"\n",
2820                stripe, kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms);
2821         if (kms > loi->loi_kms)
2822                 loi->loi_kms = kms;
2823         EXIT;
2824 }
2825 EXPORT_SYMBOL(lov_increase_kms);
2826
2827 struct obd_ops lov_obd_ops = {
2828         .o_owner               = THIS_MODULE,
2829         .o_setup               = lov_setup,
2830         .o_cleanup             = lov_cleanup,
2831         .o_connect             = lov_connect,
2832         .o_disconnect          = lov_disconnect,
2833         .o_statfs              = lov_statfs,
2834         .o_packmd              = lov_packmd,
2835         .o_unpackmd            = lov_unpackmd,
2836         .o_create              = lov_create,
2837         .o_destroy             = lov_destroy,
2838         .o_getattr             = lov_getattr,
2839         .o_getattr_async       = lov_getattr_async,
2840         .o_setattr             = lov_setattr,
2841         .o_brw                 = lov_brw,
2842         .o_brw_async           = lov_brw_async,
2843         .o_prep_async_page     = lov_prep_async_page,
2844         .o_queue_async_io      = lov_queue_async_io,
2845         .o_set_async_flags     = lov_set_async_flags,
2846         .o_queue_group_io      = lov_queue_group_io,
2847         .o_trigger_group_io    = lov_trigger_group_io,
2848         .o_teardown_async_page = lov_teardown_async_page,
2849         .o_punch               = lov_punch,
2850         .o_sync                = lov_sync,
2851         .o_enqueue             = lov_enqueue,
2852         .o_match               = lov_match,
2853         .o_change_cbdata       = lov_change_cbdata,
2854         .o_cancel              = lov_cancel,
2855         .o_cancel_unused       = lov_cancel_unused,
2856         .o_iocontrol           = lov_iocontrol,
2857         .o_get_info            = lov_get_info,
2858         .o_set_info            = lov_set_info,
2859         .o_llog_init           = lov_llog_init,
2860         .o_llog_finish         = lov_llog_finish,
2861         .o_notify              = lov_notify,
2862 };
2863
2864 int __init lov_init(void)
2865 {
2866         struct lprocfs_static_vars lvars;
2867         int rc;
2868
2869         lprocfs_init_vars(lov, &lvars);
2870         rc = class_register_type(&lov_obd_ops, lvars.module_vars,
2871                                  OBD_LOV_DEVICENAME);
2872         RETURN(rc);
2873 }
2874
2875 #ifdef __KERNEL__
2876 static void /*__exit*/ lov_exit(void)
2877 {
2878         class_unregister_type(OBD_LOV_DEVICENAME);
2879 }
2880
2881 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2882 MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver");
2883 MODULE_LICENSE("GPL");
2884
2885 module_init(lov_init);
2886 module_exit(lov_exit);
2887 #endif