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