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, newsize);
 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                         continue;
 928                 }
 929
 930                 memcpy(&tmp, oa, sizeof(tmp));
 931                 tmp.o_id = loi->loi_id;
 932                 err = obd_destroy(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
 933                                   NULL, oti);
 934                 if (err && lov->tgts[loi->loi_ost_idx].active) {
 935                         CERROR("error: destroying objid "LPX64" subobj "
 936                                LPX64" on OST idx %d: rc = %d\n",
 937                                oa->o_id, loi->loi_id, loi->loi_ost_idx, err);
 938                         if (!rc)
 939                                 rc = err;
 940                 }
 941         }
 942         RETURN(rc);
 943 }
 944
 945 static int lov_getattr(struct obd_export *exp, struct obdo *oa,
 946                        struct lov_stripe_md *lsm)
 947 {
 948         struct obdo tmp;
 949         struct lov_obd *lov;
 950         struct lov_oinfo *loi;
 951         int i, rc = 0, set = 0;
 952         ENTRY;
 953
 954         if (lsm_bad_magic(lsm))
 955                 RETURN(-EINVAL);
 956
 957         if (!exp || !exp->exp_obd)
 958                 RETURN(-ENODEV);
 959
 960         lov = &exp->exp_obd->u.lov;
 961
 962         CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
 963                lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
 964         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
 965                 int err;
 966
 967                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
 968                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
 969                         continue;
 970                 }
 971
 972                 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
 973                        "%u\n", oa->o_id, i, loi->loi_id, loi->loi_ost_idx);
 974                 /* create data objects with "parent" OA */
 975                 memcpy(&tmp, oa, sizeof(tmp));
 976                 tmp.o_id = loi->loi_id;
 977
 978                 err = obd_getattr(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp,
 979                                   NULL);
 980                 if (err) {
 981                         if (lov->tgts[loi->loi_ost_idx].active) {
 982                                 CERROR("error: getattr objid "LPX64" subobj "
 983                                        LPX64" on OST idx %d: rc = %d\n",
 984                                        oa->o_id, loi->loi_id, loi->loi_ost_idx,
 985                                        err);
 986                                 RETURN(err);
 987                         }
 988                 } else {
 989                         lov_merge_attrs(oa, &tmp, tmp.o_valid, lsm, i, &set);
 990                 }
 991         }
 992         if (!set)
 993                 rc = -EIO;
 994         RETURN(rc);
 995 }
 996
 997 static int lov_getattr_interpret(struct ptlrpc_request_set *rqset, void *data,
 998                                  int rc)
 999 {
1000         struct lov_getattr_async_args *aa = data;
1001         struct lov_stripe_md *lsm = aa->aa_lsm;
1002         struct obdo          *oa = aa->aa_oa;
1003         struct obdo          *obdos = aa->aa_obdos;
1004         struct lov_oinfo     *loi;
1005         int                   i;
1006         int                   set = 0;
1007         ENTRY;
1008
1009         if (rc == 0) {
1010                 /* NB all stripe requests succeeded to get here */
1011
1012                 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1013                      i++, loi++) {
1014                         if (obdos[i].o_valid == 0)      /* inactive stripe */
1015                                 continue;
1016
1017                         lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1018                                         i, &set);
1019                 }
1020
1021                 if (!set) {
1022                         CERROR ("No stripes had valid attrs\n");
1023                         rc = -EIO;
1024                 }
1025         }
1026
1027         OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1028         RETURN (rc);
1029 }
1030
1031 static int lov_getattr_async(struct obd_export *exp, struct obdo *oa,
1032                               struct lov_stripe_md *lsm,
1033                               struct ptlrpc_request_set *rqset)
1034 {
1035         struct obdo *obdos;
1036         struct lov_obd *lov;
1037         struct lov_oinfo *loi;
1038         struct lov_getattr_async_args *aa;
1039         int i, rc = 0, set = 0;
1040         ENTRY;
1041
1042         if (lsm_bad_magic(lsm))
1043                 RETURN(-EINVAL);
1044
1045         if (!exp || !exp->exp_obd)
1046                 RETURN(-ENODEV);
1047
1048         lov = &exp->exp_obd->u.lov;
1049
1050         OBD_ALLOC (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1051         if (obdos == NULL)
1052                 RETURN(-ENOMEM);
1053
1054         CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1055                lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size);
1056         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1057                 int err;
1058
1059                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1060                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1061                         /* leaves obdos[i].obd_valid unset */
1062                         continue;
1063                 }
1064
1065                 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
1066                        "%u\n", oa->o_id, i, loi->loi_id, loi->loi_ost_idx);
1067                 /* create data objects with "parent" OA */
1068                 memcpy(&obdos[i], oa, sizeof(obdos[i]));
1069                 obdos[i].o_id = loi->loi_id;
1070
1071                 err = obd_getattr_async(lov->tgts[loi->loi_ost_idx].ltd_exp,
1072                                          &obdos[i], NULL, rqset);
1073                 if (err) {
1074                         CERROR("error: getattr objid "LPX64" subobj "
1075                                LPX64" on OST idx %d: rc = %d\n",
1076                                oa->o_id, loi->loi_id, loi->loi_ost_idx,
1077                                err);
1078                         GOTO(out_obdos, rc = err);
1079                 }
1080                 set = 1;
1081         }
1082         if (!set)
1083                 GOTO (out_obdos, rc = -EIO);
1084
1085         LASSERT (rqset->set_interpret == NULL);
1086         rqset->set_interpret = lov_getattr_interpret;
1087         LASSERT (sizeof (rqset->set_args) >= sizeof (*aa));
1088         aa = (struct lov_getattr_async_args *)&rqset->set_args;
1089         aa->aa_lsm = lsm;
1090         aa->aa_oa = oa;
1091         aa->aa_obdos = obdos;
1092         aa->aa_lov = lov;
1093         GOTO(out, rc = 0);
1094
1095 out_obdos:
1096         OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos));
1097 out:
1098         RETURN(rc);
1099 }
1100
1101
1102 static int lov_setattr(struct obd_export *exp, struct obdo *src_oa,
1103                        struct lov_stripe_md *lsm, struct obd_trans_info *oti)
1104 {
1105         struct obdo *tmp_oa, *ret_oa;
1106         struct lov_obd *lov;
1107         struct lov_oinfo *loi;
1108         int rc = 0, i, set = 0;
1109         ENTRY;
1110
1111         if (lsm_bad_magic(lsm))
1112                 RETURN(-EINVAL);
1113
1114         if (!exp || !exp->exp_obd)
1115                 RETURN(-ENODEV);
1116
1117         /* for now, we only expect time updates here */
1118         LASSERT(!(src_oa->o_valid & ~(OBD_MD_FLID|OBD_MD_FLTYPE | OBD_MD_FLMODE|
1119                                       OBD_MD_FLATIME | OBD_MD_FLMTIME |
1120                                       OBD_MD_FLCTIME | OBD_MD_FLFLAGS |
1121                                       OBD_MD_FLSIZE)));
1122         ret_oa = obdo_alloc();
1123         if (!ret_oa)
1124                 RETURN(-ENOMEM);
1125
1126         tmp_oa = obdo_alloc();
1127         if (!tmp_oa)
1128                 GOTO(out_oa, rc = -ENOMEM);
1129
1130         lov = &exp->exp_obd->u.lov;
1131         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1132                 int err;
1133
1134                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1135                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1136                         continue;
1137                 }
1138
1139                 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1140                 tmp_oa->o_id = loi->loi_id;
1141
1142                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1143                         if (lov_stripe_offset(lsm, src_oa->o_size, i,
1144                                               &tmp_oa->o_size) < 0 &&
1145                             tmp_oa->o_size)
1146                                 tmp_oa->o_size--;
1147
1148                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1149                                i, tmp_oa->o_size, src_oa->o_size);
1150                 }
1151
1152                 err = obd_setattr(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp_oa,
1153                                   NULL, NULL);
1154                 if (err) {
1155                         if (lov->tgts[loi->loi_ost_idx].active) {
1156                                 CERROR("error: setattr objid "LPX64" subobj "
1157                                        LPX64" on OST idx %d: rc = %d\n",
1158                                        src_oa->o_id, loi->loi_id,
1159                                        loi->loi_ost_idx, err);
1160                                 if (!rc)
1161                                         rc = err;
1162                         }
1163                         continue;
1164                 }
1165                 lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm, i, &set);
1166         }
1167         if (!set && !rc)
1168                 rc = -EIO;
1169
1170         ret_oa->o_id = src_oa->o_id;
1171         memcpy(src_oa, ret_oa, sizeof(*src_oa));
1172         GOTO(out_tmp, rc);
1173 out_tmp:
1174         obdo_free(tmp_oa);
1175 out_oa:
1176         obdo_free(ret_oa);
1177         return rc;
1178 }
1179
1180 /* we have an offset in file backed by an lov and want to find out where
1181  * that offset lands in our given stripe of the file.  for the easy
1182  * case where the offset is within the stripe, we just have to scale the
1183  * offset down to make it relative to the stripe instead of the lov.
1184  *
1185  * the harder case is what to do when the offset doesn't intersect the
1186  * stripe.  callers will want start offsets clamped ahead to the start
1187  * of the nearest stripe in the file.  end offsets similarly clamped to the
1188  * nearest ending byte of a stripe in the file:
1189  *
1190  * all this function does is move offsets to the nearest region of the
1191  * stripe, and it does its work "mod" the full length of all the stripes.
1192  * consider a file with 3 stripes:
1193  *
1194  *             S                                              E
1195  * ---------------------------------------------------------------------
1196  * |    0    |     1     |     2     |    0    |     1     |     2     |
1197  * ---------------------------------------------------------------------
1198  *
1199  * to find stripe 1's offsets for S and E, it divides by the full stripe
1200  * width and does its math in the context of a single set of stripes:
1201  *
1202  *             S         E
1203  * -----------------------------------
1204  * |    0    |     1     |     2     |
1205  * -----------------------------------
1206  *
1207  * it'll notice that E is outside stripe 1 and clamp it to the end of the
1208  * stripe, then multiply it back out by lov_off to give the real offsets in
1209  * the stripe:
1210  *
1211  *   S                   E
1212  * ---------------------------------------------------------------------
1213  * |    1    |     1     |     1     |    1    |     1     |     1     |
1214  * ---------------------------------------------------------------------
1215  *
1216  * it would have done similarly and pulled S forward to the start of a 1
1217  * stripe if, say, S had landed in a 0 stripe.
1218  *
1219  * this rounding isn't always correct.  consider an E lov offset that lands
1220  * on a 0 stripe, the "mod stripe width" math will pull it forward to the
1221  * start of a 1 stripe, when in fact it wanted to be rounded back to the end
1222  * of a previous 1 stripe.  this logic is handled by callers and this is why:
1223  *
1224  * this function returns < 0 when the offset was "before" the stripe and
1225  * was moved forward to the start of the stripe in question;  0 when it
1226  * falls in the stripe and no shifting was done; > 0 when the offset
1227  * was outside the stripe and was pulled back to its final byte. */
1228 static int lov_stripe_offset(struct lov_stripe_md *lsm, obd_off lov_off,
1229                              int stripeno, obd_off *obd_off)
1230 {
1231         unsigned long ssize  = lsm->lsm_stripe_size;
1232         unsigned long swidth = ssize * lsm->lsm_stripe_count;
1233         unsigned long stripe_off, this_stripe;
1234         int ret = 0;
1235
1236         if (lov_off == OBD_OBJECT_EOF) {
1237                 *obd_off = OBD_OBJECT_EOF;
1238                 return 0;
1239         }
1240
1241         /* do_div(a, b) returns a % b, and a = a / b */
1242         stripe_off = do_div(lov_off, swidth);
1243
1244         this_stripe = stripeno * ssize;
1245         if (stripe_off < this_stripe) {
1246                 stripe_off = 0;
1247                 ret = -1;
1248         } else {
1249                 stripe_off -= this_stripe;
1250
1251                 if (stripe_off >= ssize) {
1252                         stripe_off = ssize;
1253                         ret = 1;
1254                 }
1255         }
1256
1257         *obd_off = lov_off * ssize + stripe_off;
1258         return ret;
1259 }
1260
1261 /* Given a whole-file size and a stripe number, give the file size which
1262  * corresponds to the individual object of that stripe.
1263  *
1264  * This behaves basically in the same was as lov_stripe_offset, except that
1265  * file sizes falling before the beginning of a stripe are clamped to the end
1266  * of the previous stripe, not the beginning of the next:
1267  *
1268  *                                               S
1269  * ---------------------------------------------------------------------
1270  * |    0    |     1     |     2     |    0    |     1     |     2     |
1271  * ---------------------------------------------------------------------
1272  *
1273  * if clamped to stripe 2 becomes:
1274  *
1275  *                                   S
1276  * ---------------------------------------------------------------------
1277  * |    0    |     1     |     2     |    0    |     1     |     2     |
1278  * ---------------------------------------------------------------------
1279  */
1280 static obd_off lov_size_to_stripe(struct lov_stripe_md *lsm, obd_off file_size,
1281                                   int stripeno)
1282 {
1283         unsigned long ssize  = lsm->lsm_stripe_size;
1284         unsigned long swidth = ssize * lsm->lsm_stripe_count;
1285         unsigned long stripe_off, this_stripe;
1286
1287         if (file_size == OBD_OBJECT_EOF)
1288                 return OBD_OBJECT_EOF;
1289
1290         /* do_div(a, b) returns a % b, and a = a / b */
1291         stripe_off = do_div(file_size, swidth);
1292
1293         this_stripe = stripeno * ssize;
1294         if (stripe_off < this_stripe) {
1295                 /* Move to end of previous stripe, or zero */
1296                 if (file_size > 0) {
1297                         file_size--;
1298                         stripe_off = ssize;
1299                 } else {
1300                         stripe_off = 0;
1301                 }
1302         } else {
1303                 stripe_off -= this_stripe;
1304
1305                 if (stripe_off >= ssize) {
1306                         /* Clamp to end of this stripe */
1307                         stripe_off = ssize;
1308                 }
1309         }
1310
1311         return (file_size * ssize + stripe_off);
1312 }
1313
1314 /* given an extent in an lov and a stripe, calculate the extent of the stripe
1315  * that is contained within the lov extent.  this returns true if the given
1316  * stripe does intersect with the lov extent. */
1317 static int lov_stripe_intersects(struct lov_stripe_md *lsm, int stripeno,
1318                                  obd_off start, obd_off end,
1319                                  obd_off *obd_start, obd_off *obd_end)
1320 {
1321         int start_side, end_side;
1322
1323         start_side = lov_stripe_offset(lsm, start, stripeno, obd_start);
1324         end_side = lov_stripe_offset(lsm, end, stripeno, obd_end);
1325
1326         CDEBUG(D_INODE, "["LPU64"->"LPU64"] -> [(%d) "LPU64"->"LPU64" (%d)]\n",
1327                start, end, start_side, *obd_start, *obd_end, end_side);
1328
1329         /* this stripe doesn't intersect the file extent when neither
1330          * start or the end intersected the stripe and obd_start and
1331          * obd_end got rounded up to the save value. */
1332         if (start_side != 0 && end_side != 0 && *obd_start == *obd_end)
1333                 return 0;
1334
1335         /* as mentioned in the lov_stripe_offset commentary, end
1336          * might have been shifted in the wrong direction.  This
1337          * happens when an end offset is before the stripe when viewed
1338          * through the "mod stripe size" math. we detect it being shifted
1339          * in the wrong direction and touch it up.
1340          * interestingly, this can't underflow since end must be > start
1341          * if we passed through the previous check.
1342          * (should we assert for that somewhere?) */
1343         if (end_side != 0)
1344                 (*obd_end)--;
1345
1346         return 1;
1347 }
1348
1349 /* compute which stripe number "lov_off" will be written into */
1350 static int lov_stripe_number(struct lov_stripe_md *lsm, obd_off lov_off)
1351 {
1352         unsigned long ssize  = lsm->lsm_stripe_size;
1353         unsigned long swidth = ssize * lsm->lsm_stripe_count;
1354         unsigned long stripe_off;
1355
1356         stripe_off = do_div(lov_off, swidth);
1357
1358         return stripe_off / ssize;
1359 }
1360
1361 /* FIXME: maybe we'll just make one node the authoritative attribute node, then
1362  * we can send this 'punch' to just the authoritative node and the nodes
1363  * that the punch will affect. */
1364 static int lov_punch(struct obd_export *exp, struct obdo *oa,
1365                      struct lov_stripe_md *lsm,
1366                      obd_off start, obd_off end, struct obd_trans_info *oti)
1367 {
1368         struct obdo tmp;
1369         struct lov_obd *lov;
1370         struct lov_oinfo *loi;
1371         int rc = 0, i;
1372         ENTRY;
1373
1374         if (lsm_bad_magic(lsm))
1375                 RETURN(-EINVAL);
1376
1377         if (!exp || !exp->exp_obd)
1378                 RETURN(-ENODEV);
1379
1380         lov = &exp->exp_obd->u.lov;
1381         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1382                 obd_off starti, endi;
1383                 int err;
1384
1385                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1386                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1387                         continue;
1388                 }
1389
1390                 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1391                         continue;
1392
1393                 /* create data objects with "parent" OA */
1394                 memcpy(&tmp, oa, sizeof(tmp));
1395                 tmp.o_id = loi->loi_id;
1396
1397                 err = obd_punch(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp, NULL,
1398                                 starti, endi, NULL);
1399                 if (err) {
1400                         if (lov->tgts[loi->loi_ost_idx].active) {
1401                                 CERROR("error: punch objid "LPX64" subobj "LPX64
1402                                        " on OST idx %d: rc = %d\n", oa->o_id,
1403                                        loi->loi_id, loi->loi_ost_idx, err);
1404                         }
1405                         if (!rc)
1406                                 rc = err;
1407                 } else {
1408                         loi->loi_kms = loi->loi_rss = starti;
1409                 }
1410         }
1411         RETURN(rc);
1412 }
1413
1414 static int lov_sync(struct obd_export *exp, struct obdo *oa,
1415                     struct lov_stripe_md *lsm, obd_off start, obd_off end)
1416 {
1417         struct obdo *tmp;
1418         struct lov_obd *lov;
1419         struct lov_oinfo *loi;
1420         int rc = 0, i;
1421         ENTRY;
1422
1423         if (lsm_bad_magic(lsm))
1424                 RETURN(-EINVAL);
1425
1426         if (!exp->exp_obd)
1427                 RETURN(-ENODEV);
1428
1429         tmp = obdo_alloc();
1430         if (!tmp)
1431                 RETURN(-ENOMEM);
1432
1433         lov = &exp->exp_obd->u.lov;
1434         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
1435                 obd_off starti, endi;
1436                 int err;
1437
1438                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
1439                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1440                         continue;
1441                 }
1442
1443                 if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi))
1444                         continue;
1445
1446                 memcpy(tmp, oa, sizeof(*tmp));
1447                 tmp->o_id = loi->loi_id;
1448
1449                 err = obd_sync(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp, NULL,
1450                                starti, endi);
1451                 if (err) {
1452                         if (lov->tgts[loi->loi_ost_idx].active) {
1453                                 CERROR("error: fsync objid "LPX64" subobj "LPX64
1454                                        " on OST idx %d: rc = %d\n", oa->o_id,
1455                                        loi->loi_id, loi->loi_ost_idx, err);
1456                         }
1457                         if (!rc)
1458                                 rc = err;
1459                 }
1460         }
1461
1462         obdo_free(tmp);
1463         RETURN(rc);
1464 }
1465
1466 static int lov_brw_check(struct lov_obd *lov, struct obdo *oa,
1467                          struct lov_stripe_md *lsm,
1468                          obd_count oa_bufs, struct brw_page *pga)
1469 {
1470         int i, rc = 0;
1471
1472         /* The caller just wants to know if there's a chance that this
1473          * I/O can succeed */
1474         for (i = 0; i < oa_bufs; i++) {
1475                 int stripe = lov_stripe_number(lsm, pga[i].off);
1476                 int ost = lsm->lsm_oinfo[stripe].loi_ost_idx;
1477                 struct ldlm_extent ext, subext;
1478                 ext.start = pga[i].off;
1479                 ext.end = pga[i].off + pga[i].count;
1480
1481                 if (!lov_stripe_intersects(lsm, i, ext.start, ext.end,
1482                                            &subext.start, &subext.end))
1483                         continue;
1484
1485                 if (lov->tgts[ost].active == 0) {
1486                         CDEBUG(D_HA, "lov idx %d inactive\n", ost);
1487                         return -EIO;
1488                 }
1489                 rc = obd_brw(OBD_BRW_CHECK, lov->tgts[stripe].ltd_exp, oa,
1490                              NULL, 1, &pga[i], NULL);
1491                 if (rc)
1492                         break;
1493         }
1494         return rc;
1495 }
1496
1497 static int lov_brw(int cmd, struct obd_export *exp, struct obdo *src_oa,
1498                    struct lov_stripe_md *lsm, obd_count oa_bufs,
1499                    struct brw_page *pga, struct obd_trans_info *oti)
1500 {
1501         struct {
1502                 int bufct;
1503                 int index;
1504                 int subcount;
1505                 struct lov_stripe_md lsm;
1506                 int ost_idx;
1507         } *stripeinfo, *si, *si_last;
1508         struct obdo *ret_oa = NULL, *tmp_oa = NULL;
1509         struct lov_obd *lov;
1510         struct brw_page *ioarr;
1511         struct lov_oinfo *loi;
1512         int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count, set = 0;
1513         ENTRY;
1514
1515         if (lsm_bad_magic(lsm))
1516                 RETURN(-EINVAL);
1517
1518         lov = &exp->exp_obd->u.lov;
1519
1520         if (cmd == OBD_BRW_CHECK) {
1521                 rc = lov_brw_check(lov, src_oa, lsm, oa_bufs, pga);
1522                 RETURN(rc);
1523         }
1524
1525         OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
1526         if (!stripeinfo)
1527                 RETURN(-ENOMEM);
1528
1529         OBD_ALLOC(where, sizeof(*where) * oa_bufs);
1530         if (!where)
1531                 GOTO(out_sinfo, rc = -ENOMEM);
1532
1533         OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
1534         if (!ioarr)
1535                 GOTO(out_where, rc = -ENOMEM);
1536
1537         if (src_oa) {
1538                 ret_oa = obdo_alloc();
1539                 if (!ret_oa)
1540                         GOTO(out_ioarr, rc = -ENOMEM);
1541
1542                 tmp_oa = obdo_alloc();
1543                 if (!tmp_oa)
1544                         GOTO(out_oa, rc = -ENOMEM);
1545         }
1546
1547         for (i = 0; i < oa_bufs; i++) {
1548                 where[i] = lov_stripe_number(lsm, pga[i].off);
1549                 stripeinfo[where[i]].bufct++;
1550         }
1551
1552         for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
1553              i < stripe_count; i++, loi++, si_last = si, si++) {
1554                 if (i > 0)
1555                         si->index = si_last->index + si_last->bufct;
1556                 si->lsm.lsm_object_id = loi->loi_id;
1557                 si->ost_idx = loi->loi_ost_idx;
1558         }
1559
1560         for (i = 0; i < oa_bufs; i++) {
1561                 int which = where[i];
1562                 int shift;
1563
1564                 shift = stripeinfo[which].index + stripeinfo[which].subcount;
1565                 LASSERT(shift < oa_bufs);
1566                 ioarr[shift] = pga[i];
1567                 lov_stripe_offset(lsm, pga[i].off, which, &ioarr[shift].off);
1568                 stripeinfo[which].subcount++;
1569         }
1570
1571         for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
1572                 int shift = si->index;
1573
1574                 if (lov->tgts[si->ost_idx].active == 0) {
1575                         CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
1576                         GOTO(out_oa, rc = -EIO);
1577                 }
1578
1579                 if (si->bufct) {
1580                         LASSERT(shift < oa_bufs);
1581                         if (src_oa)
1582                                 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
1583
1584                         tmp_oa->o_id = si->lsm.lsm_object_id;
1585                         rc = obd_brw(cmd, lov->tgts[si->ost_idx].ltd_exp,
1586                                      tmp_oa, &si->lsm, si->bufct,
1587                                      &ioarr[shift], oti);
1588                         if (rc)
1589                                 GOTO(out_ioarr, rc);
1590
1591                         lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm,
1592                                         i, &set);
1593                 }
1594         }
1595
1596         ret_oa->o_id = src_oa->o_id;
1597         memcpy(src_oa, ret_oa, sizeof(*src_oa));
1598
1599         GOTO(out_oa, rc);
1600  out_oa:
1601         if (tmp_oa)
1602                 obdo_free(tmp_oa);
1603         if (ret_oa)
1604                 obdo_free(ret_oa);
1605  out_ioarr:
1606         OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1607  out_where:
1608         OBD_FREE(where, sizeof(*where) * oa_bufs);
1609  out_sinfo:
1610         OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
1611         return rc;
1612 }
1613
1614 static int lov_brw_interpret(struct ptlrpc_request_set *reqset, void *data,
1615                              int rc)
1616 {
1617         struct lov_brw_async_args *aa = data;
1618         struct lov_stripe_md *lsm = aa->aa_lsm;
1619         obd_count             oa_bufs = aa->aa_oa_bufs;
1620         struct obdo          *oa = aa->aa_oa;
1621         struct obdo          *obdos = aa->aa_obdos;
1622         struct brw_page      *ioarr = aa->aa_ioarr;
1623         struct lov_oinfo     *loi;
1624         int i, set = 0;
1625         ENTRY;
1626
1627         if (rc == 0) {
1628                 /* NB all stripe requests succeeded to get here */
1629
1630                 for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1631                      i++, loi++) {
1632                         if (obdos[i].o_valid == 0)      /* inactive stripe */
1633                                 continue;
1634
1635                         lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm,
1636                                         i, &set);
1637                 }
1638
1639                 if (!set) {
1640                         CERROR("No stripes had valid attrs\n");
1641                         rc = -EIO;
1642                 }
1643         }
1644         oa->o_id = lsm->lsm_object_id;
1645
1646         OBD_FREE(obdos, lsm->lsm_stripe_count * sizeof(*obdos));
1647         OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1648         RETURN(rc);
1649 }
1650
1651 static int lov_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1652                          struct lov_stripe_md *lsm, obd_count oa_bufs,
1653                          struct brw_page *pga, struct ptlrpc_request_set *set,
1654                          struct obd_trans_info *oti)
1655 {
1656         struct {
1657                 int bufct;
1658                 int index;
1659                 int subcount;
1660                 struct lov_stripe_md lsm;
1661                 int ost_idx;
1662         } *stripeinfo, *si, *si_last;
1663         struct lov_obd *lov;
1664         struct brw_page *ioarr;
1665         struct obdo *obdos = NULL;
1666         struct lov_oinfo *loi;
1667         struct lov_brw_async_args *aa;
1668         int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count;
1669         ENTRY;
1670
1671         if (lsm_bad_magic(lsm))
1672                 RETURN(-EINVAL);
1673
1674         lov = &exp->exp_obd->u.lov;
1675
1676         if (cmd == OBD_BRW_CHECK) {
1677                 rc = lov_brw_check(lov, oa, lsm, oa_bufs, pga);
1678                 RETURN(rc);
1679         }
1680
1681         OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo));
1682         if (!stripeinfo)
1683                 RETURN(-ENOMEM);
1684
1685         OBD_ALLOC(where, sizeof(*where) * oa_bufs);
1686         if (!where)
1687                 GOTO(out_sinfo, rc = -ENOMEM);
1688
1689         if (oa) {
1690                 OBD_ALLOC(obdos, sizeof(*obdos) * stripe_count);
1691                 if (!obdos)
1692                         GOTO(out_where, rc = -ENOMEM);
1693         }
1694
1695         OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs);
1696         if (!ioarr)
1697                 GOTO(out_obdos, rc = -ENOMEM);
1698
1699         for (i = 0; i < oa_bufs; i++) {
1700                 where[i] = lov_stripe_number(lsm, pga[i].off);
1701                 stripeinfo[where[i]].bufct++;
1702         }
1703
1704         for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo;
1705              i < stripe_count; i++, loi++, si_last = si, si++) {
1706                 if (i > 0)
1707                         si->index = si_last->index + si_last->bufct;
1708                 si->lsm.lsm_object_id = loi->loi_id;
1709                 si->ost_idx = loi->loi_ost_idx;
1710
1711                 if (oa) {
1712                         memcpy(&obdos[i], oa, sizeof(*obdos));
1713                         obdos[i].o_id = si->lsm.lsm_object_id;
1714                 }
1715         }
1716
1717         for (i = 0; i < oa_bufs; i++) {
1718                 int which = where[i];
1719                 int shift;
1720
1721                 shift = stripeinfo[which].index + stripeinfo[which].subcount;
1722                 LASSERT(shift < oa_bufs);
1723                 ioarr[shift] = pga[i];
1724                 lov_stripe_offset(lsm, pga[i].off, which, &ioarr[shift].off);
1725                 stripeinfo[which].subcount++;
1726         }
1727
1728         for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) {
1729                 int shift = si->index;
1730
1731                 if (si->bufct == 0)
1732                         continue;
1733
1734                 if (lov->tgts[si->ost_idx].active == 0) {
1735                         CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx);
1736                         GOTO(out_ioarr, rc = -EIO);
1737                 }
1738
1739                 LASSERT(shift < oa_bufs);
1740
1741                 rc = obd_brw_async(cmd, lov->tgts[si->ost_idx].ltd_exp,
1742                                    &obdos[i], &si->lsm, si->bufct,
1743                                    &ioarr[shift], set, oti);
1744                 if (rc)
1745                         GOTO(out_ioarr, rc);
1746         }
1747         LASSERT(rc == 0);
1748         LASSERT(set->set_interpret == NULL);
1749         set->set_interpret = (set_interpreter_func)lov_brw_interpret;
1750         LASSERT(sizeof(set->set_args) >= sizeof(struct lov_brw_async_args));
1751         aa = (struct lov_brw_async_args *)&set->set_args;
1752         aa->aa_lsm = lsm;
1753         aa->aa_obdos = obdos;
1754         aa->aa_oa = oa;
1755         aa->aa_ioarr = ioarr;
1756         aa->aa_oa_bufs = oa_bufs;
1757
1758         /* Don't free ioarr or obdos - that's done in lov_brw_interpret */
1759         GOTO(out_where, rc);
1760
1761  out_ioarr:
1762         OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs);
1763  out_obdos:
1764         OBD_FREE(obdos, stripe_count * sizeof(*obdos));
1765  out_where:
1766         OBD_FREE(where, sizeof(*where) * oa_bufs);
1767  out_sinfo:
1768         OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo));
1769         return rc;
1770 }
1771
1772 struct lov_async_page *lap_from_cookie(void *cookie)
1773 {
1774         struct lov_async_page *lap = cookie;
1775         if (lap->lap_magic != LAP_MAGIC)
1776                 return ERR_PTR(-EINVAL);
1777         return lap;
1778 };
1779
1780 static int lov_ap_make_ready(void *data, int cmd)
1781 {
1782         struct lov_async_page *lap = lap_from_cookie(data);
1783         /* XXX should these assert? */
1784         if (IS_ERR(lap))
1785                 return -EINVAL;
1786
1787         return lap->lap_caller_ops->ap_make_ready(lap->lap_caller_data, cmd);
1788 }
1789 static int lov_ap_refresh_count(void *data, int cmd)
1790 {
1791         struct lov_async_page *lap = lap_from_cookie(data);
1792         if (IS_ERR(lap))
1793                 return -EINVAL;
1794
1795         return lap->lap_caller_ops->ap_refresh_count(lap->lap_caller_data,
1796                                                      cmd);
1797 }
1798 static void lov_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
1799 {
1800         struct lov_async_page *lap = lap_from_cookie(data);
1801         /* XXX should these assert? */
1802         if (IS_ERR(lap))
1803                 return;
1804
1805         lap->lap_caller_ops->ap_fill_obdo(lap->lap_caller_data, cmd, oa);
1806         /* XXX woah, shouldn't we be altering more here?  size? */
1807         oa->o_id = lap->lap_loi_id;
1808 }
1809 static void lov_ap_completion(void *data, int cmd, int rc)
1810 {
1811         struct lov_async_page *lap = lap_from_cookie(data);
1812         if (IS_ERR(lap))
1813                 return;
1814
1815         /* in a raid1 regime this would down a count of many ios
1816          * in flight, onl calling the caller_ops completion when all
1817          * the raid1 ios are complete */
1818         lap->lap_caller_ops->ap_completion(lap->lap_caller_data, cmd, rc);
1819 }
1820
1821 static struct obd_async_page_ops lov_async_page_ops = {
1822         .ap_make_ready =        lov_ap_make_ready,
1823         .ap_refresh_count =     lov_ap_refresh_count,
1824         .ap_fill_obdo =         lov_ap_fill_obdo,
1825         .ap_completion =        lov_ap_completion,
1826 };
1827
1828 int lov_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1829                            struct lov_oinfo *loi, struct page *page,
1830                            obd_off offset, struct obd_async_page_ops *ops,
1831                            void *data, void **res)
1832 {
1833         struct lov_obd *lov = &exp->exp_obd->u.lov;
1834         struct lov_async_page *lap;
1835         int rc;
1836         ENTRY;
1837
1838         if (lsm_bad_magic(lsm))
1839                 RETURN(-EINVAL);
1840         LASSERT(loi == NULL);
1841
1842         OBD_ALLOC(lap, sizeof(*lap));
1843         if (lap == NULL)
1844                 RETURN(-ENOMEM);
1845
1846         lap->lap_magic = LAP_MAGIC;
1847         lap->lap_caller_ops = ops;
1848         lap->lap_caller_data = data;
1849
1850         /* for now only raid 0 which passes through */
1851         lap->lap_stripe = lov_stripe_number(lsm, offset);
1852         lov_stripe_offset(lsm, offset, lap->lap_stripe, &lap->lap_sub_offset);
1853         loi = &lsm->lsm_oinfo[lap->lap_stripe];
1854
1855         /* so the callback doesn't need the lsm */
1856         lap->lap_loi_id = loi->loi_id;
1857
1858         rc = obd_prep_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
1859                                  lsm, loi, page, lap->lap_sub_offset,
1860                                  &lov_async_page_ops, lap,
1861                                  &lap->lap_sub_cookie);
1862         if (rc) {
1863                 OBD_FREE(lap, sizeof(*lap));
1864                 RETURN(rc);
1865         }
1866         CDEBUG(D_CACHE, "lap %p page %p cookie %p off "LPU64"\n", lap, page,
1867                lap->lap_sub_cookie, offset);
1868         *res = lap;
1869         RETURN(0);
1870 }
1871
1872 static int lov_queue_async_io(struct obd_export *exp,
1873                               struct lov_stripe_md *lsm,
1874                               struct lov_oinfo *loi, void *cookie,
1875                               int cmd, obd_off off, int count,
1876                               obd_flag brw_flags, obd_flag async_flags)
1877 {
1878         struct lov_obd *lov = &exp->exp_obd->u.lov;
1879         struct lov_async_page *lap;
1880         int rc;
1881
1882         LASSERT(loi == NULL);
1883
1884         if (lsm_bad_magic(lsm))
1885                 RETURN(-EINVAL);
1886
1887         lap = lap_from_cookie(cookie);
1888         if (IS_ERR(lap))
1889                 RETURN(PTR_ERR(lap));
1890
1891         loi = &lsm->lsm_oinfo[lap->lap_stripe];
1892         rc = obd_queue_async_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm,
1893                                 loi, lap->lap_sub_cookie, cmd, off, count,
1894                                 brw_flags, async_flags);
1895         RETURN(rc);
1896 }
1897
1898 static int lov_set_async_flags(struct obd_export *exp,
1899                                struct lov_stripe_md *lsm,
1900                                struct lov_oinfo *loi, void *cookie,
1901                                obd_flag async_flags)
1902 {
1903         struct lov_obd *lov = &exp->exp_obd->u.lov;
1904         struct lov_async_page *lap;
1905         int rc;
1906
1907         LASSERT(loi == NULL);
1908
1909         if (lsm_bad_magic(lsm))
1910                 RETURN(-EINVAL);
1911
1912         lap = lap_from_cookie(cookie);
1913         if (IS_ERR(lap))
1914                 RETURN(PTR_ERR(lap));
1915
1916         loi = &lsm->lsm_oinfo[lap->lap_stripe];
1917         rc = obd_set_async_flags(lov->tgts[loi->loi_ost_idx].ltd_exp,
1918                                  lsm, loi, lap->lap_sub_cookie, async_flags);
1919         RETURN(rc);
1920 }
1921
1922 static int lov_queue_group_io(struct obd_export *exp,
1923                               struct lov_stripe_md *lsm,
1924                               struct lov_oinfo *loi,
1925                               struct obd_io_group *oig, void *cookie,
1926                               int cmd, obd_off off, int count,
1927                               obd_flag brw_flags, obd_flag async_flags)
1928 {
1929         struct lov_obd *lov = &exp->exp_obd->u.lov;
1930         struct lov_async_page *lap;
1931         int rc;
1932
1933         LASSERT(loi == NULL);
1934
1935         if (lsm_bad_magic(lsm))
1936                 RETURN(-EINVAL);
1937
1938         lap = lap_from_cookie(cookie);
1939         if (IS_ERR(lap))
1940                 RETURN(PTR_ERR(lap));
1941
1942         loi = &lsm->lsm_oinfo[lap->lap_stripe];
1943         rc = obd_queue_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm, loi,
1944                                 oig, lap->lap_sub_cookie, cmd, off, count,
1945                                 brw_flags, async_flags);
1946         RETURN(rc);
1947 }
1948
1949 /* this isn't exactly optimal.  we may have queued sync io in oscs on
1950  * all stripes, but we don't record that fact at queue time.  so we
1951  * trigger sync io on all stripes. */
1952 static int lov_trigger_group_io(struct obd_export *exp,
1953                                 struct lov_stripe_md *lsm,
1954                                 struct lov_oinfo *loi,
1955                                 struct obd_io_group *oig)
1956 {
1957         struct lov_obd *lov = &exp->exp_obd->u.lov;
1958         int rc = 0, i, err;
1959
1960         LASSERT(loi == NULL);
1961
1962         if (lsm_bad_magic(lsm))
1963                 RETURN(-EINVAL);
1964
1965         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
1966              i++, loi++) {
1967                 err = obd_trigger_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp,
1968                                            lsm, loi, oig);
1969                 if (rc == 0 && err != 0)
1970                         rc = err;
1971         };
1972         RETURN(rc);
1973 }
1974
1975 static int lov_teardown_async_page(struct obd_export *exp,
1976                                    struct lov_stripe_md *lsm,
1977                                    struct lov_oinfo *loi, void *cookie)
1978 {
1979         struct lov_obd *lov = &exp->exp_obd->u.lov;
1980         struct lov_async_page *lap;
1981         int rc;
1982
1983         LASSERT(loi == NULL);
1984
1985         if (lsm_bad_magic(lsm))
1986                 RETURN(-EINVAL);
1987
1988         lap = lap_from_cookie(cookie);
1989         if (IS_ERR(lap))
1990                 RETURN(PTR_ERR(lap));
1991
1992         loi = &lsm->lsm_oinfo[lap->lap_stripe];
1993         rc = obd_teardown_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp,
1994                                      lsm, loi, lap->lap_sub_cookie);
1995         if (rc) {
1996                 CERROR("unable to teardown sub cookie %p: %d\n",
1997                        lap->lap_sub_cookie, rc);
1998                 RETURN(rc);
1999         }
2000         OBD_FREE(lap, sizeof(*lap));
2001         RETURN(rc);
2002 }
2003
2004 static int lov_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2005                        __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2006                        int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2007                        void *data,__u32 lvb_len, void *lvb_swabber,
2008                        struct lustre_handle *lockh)
2009 {
2010         struct lov_lock_handles *lov_lockh = NULL;
2011         struct lustre_handle *lov_lockhp;
2012         struct lov_obd *lov;
2013         struct lov_oinfo *loi;
2014         char submd_buf[sizeof(struct lov_stripe_md) + sizeof(struct lov_oinfo)];
2015         struct lov_stripe_md *submd = (void *)submd_buf;
2016         ldlm_error_t rc;
2017         int i, save_flags = *flags;
2018         ENTRY;
2019
2020         if (lsm_bad_magic(lsm))
2021                 RETURN(-EINVAL);
2022
2023         /* we should never be asked to replay a lock this way. */
2024         LASSERT((*flags & LDLM_FL_REPLAY) == 0);
2025
2026         if (!exp || !exp->exp_obd)
2027                 RETURN(-ENODEV);
2028
2029         if (lsm->lsm_stripe_count > 1) {
2030                 lov_lockh = lov_llh_new(lsm);
2031                 if (lov_lockh == NULL)
2032                         RETURN(-ENOMEM);
2033
2034                 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2035                 lov_lockhp = lov_lockh->llh_handles;
2036         } else {
2037                 lov_lockhp = lockh;
2038         }
2039
2040         lov = &exp->exp_obd->u.lov;
2041         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2042              i++, loi++, lov_lockhp++) {
2043                 ldlm_policy_data_t sub_ext;
2044
2045                 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2046                                            policy->l_extent.end,
2047                                            &sub_ext.l_extent.start,
2048                                            &sub_ext.l_extent.end))
2049                         continue;
2050
2051                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2052                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2053                         continue;
2054                 }
2055
2056                 /* XXX LOV STACKING: submd should be from the subobj */
2057                 submd->lsm_object_id = loi->loi_id;
2058                 submd->lsm_stripe_count = 0;
2059                 submd->lsm_oinfo->loi_kms_valid = loi->loi_kms_valid;
2060                 submd->lsm_oinfo->loi_rss = loi->loi_rss;
2061                 submd->lsm_oinfo->loi_kms = loi->loi_kms;
2062                 loi->loi_mtime = submd->lsm_oinfo->loi_mtime;
2063                 /* XXX submd is not fully initialized here */
2064                 *flags = save_flags;
2065                 rc = obd_enqueue(lov->tgts[loi->loi_ost_idx].ltd_exp, submd,
2066                                  type, &sub_ext, mode, flags, bl_cb, cp_cb,
2067                                  gl_cb, data, lvb_len, lvb_swabber, lov_lockhp);
2068
2069                 /* XXX FIXME: This unpleasantness doesn't belong here at *all*.
2070                  * It belongs in the OSC, except that the OSC doesn't have
2071                  * access to the real LOI -- it gets a copy, that we created
2072                  * above, and that copy can be arbitrarily out of date.
2073                  *
2074                  * The LOV API is due for a serious rewriting anyways, and this
2075                  * can be addressed then. */
2076                 if (rc == ELDLM_OK) {
2077                         struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp);
2078                         __u64 tmp = submd->lsm_oinfo->loi_rss;
2079
2080                         LASSERT(lock != NULL);
2081                         loi->loi_rss = tmp;
2082                         /* Extend KMS up to the end of this lock and no further
2083                          * A lock on [x,y] means a KMS of up to y + 1 bytes! */
2084                         if (tmp > lock->l_policy_data.l_extent.end)
2085                                 tmp = lock->l_policy_data.l_extent.end + 1;
2086                         if (tmp >= loi->loi_kms) {
2087                                 CDEBUG(D_INODE, "lock acquired, setting rss="
2088                                        LPU64", kms="LPU64"\n", loi->loi_rss,
2089                                        tmp);
2090                                 loi->loi_kms = tmp;
2091                                 loi->loi_kms_valid = 1;
2092                         } else {
2093                                 CDEBUG(D_INODE, "lock acquired, setting rss="
2094                                        LPU64"; leaving kms="LPU64", end="LPU64
2095                                        "\n", loi->loi_rss, loi->loi_kms,
2096                                        lock->l_policy_data.l_extent.end);
2097                         }
2098                         ldlm_lock_allow_match(lock);
2099                         LDLM_LOCK_PUT(lock);
2100                 } else if (rc == ELDLM_LOCK_ABORTED &&
2101                            save_flags & LDLM_FL_HAS_INTENT) {
2102                         memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2103                         loi->loi_rss = submd->lsm_oinfo->loi_rss;
2104                         CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
2105                                " kms="LPU64"\n", loi->loi_rss, loi->loi_kms);
2106                 } else {
2107                         memset(lov_lockhp, 0, sizeof(*lov_lockhp));
2108                         if (lov->tgts[loi->loi_ost_idx].active) {
2109                                 CERROR("error: enqueue objid "LPX64" subobj "
2110                                        LPX64" on OST idx %d: rc = %d\n",
2111                                        lsm->lsm_object_id, loi->loi_id,
2112                                        loi->loi_ost_idx, rc);
2113                                 GOTO(out_locks, rc);
2114                         }
2115                 }
2116         }
2117         if (lsm->lsm_stripe_count > 1)
2118                 lov_llh_put(lov_lockh);
2119         RETURN(ELDLM_OK);
2120
2121  out_locks:
2122         while (loi--, lov_lockhp--, i-- > 0) {
2123                 struct lov_stripe_md submd;
2124                 int err;
2125
2126                 if (lov_lockhp->cookie == 0)
2127                         continue;
2128
2129                 /* XXX LOV STACKING: submd should be from the subobj */
2130                 submd.lsm_object_id = loi->loi_id;
2131                 submd.lsm_stripe_count = 0;
2132                 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2133                                  mode, lov_lockhp);
2134                 if (err && lov->tgts[loi->loi_ost_idx].active) {
2135                         CERROR("error: cancelling objid "LPX64" on OST "
2136                                "idx %d after enqueue error: rc = %d\n",
2137                                loi->loi_id, loi->loi_ost_idx, err);
2138                 }
2139         }
2140
2141         if (lsm->lsm_stripe_count > 1) {
2142                 lov_llh_destroy(lov_lockh);
2143                 lov_llh_put(lov_lockh);
2144         }
2145         return rc;
2146 }
2147
2148 static int lov_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2149                      __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2150                      int *flags, void *data, struct lustre_handle *lockh)
2151 {
2152         struct lov_lock_handles *lov_lockh = NULL;
2153         struct lustre_handle *lov_lockhp;
2154         struct lov_obd *lov;
2155         struct lov_oinfo *loi;
2156         struct lov_stripe_md submd;
2157         ldlm_error_t rc = 0;
2158         int i;
2159         ENTRY;
2160
2161         if (lsm_bad_magic(lsm))
2162                 RETURN(-EINVAL);
2163
2164         if (!exp || !exp->exp_obd)
2165                 RETURN(-ENODEV);
2166
2167         if (lsm->lsm_stripe_count > 1) {
2168                 lov_lockh = lov_llh_new(lsm);
2169                 if (lov_lockh == NULL)
2170                         RETURN(-ENOMEM);
2171
2172                 lockh->cookie = lov_lockh->llh_handle.h_cookie;
2173                 lov_lockhp = lov_lockh->llh_handles;
2174         } else {
2175                 lov_lockhp = lockh;
2176         }
2177
2178         lov = &exp->exp_obd->u.lov;
2179         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2180              i++, loi++, lov_lockhp++) {
2181                 ldlm_policy_data_t sub_ext;
2182                 int lov_flags;
2183
2184                 if (!lov_stripe_intersects(lsm, i, policy->l_extent.start,
2185                                            policy->l_extent.end,
2186                                            &sub_ext.l_extent.start,
2187                                            &sub_ext.l_extent.end))
2188                         continue;
2189
2190                 if (lov->tgts[loi->loi_ost_idx].active == 0) {
2191                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2192                         rc = -EIO;
2193                         break;
2194                 }
2195
2196                 /* XXX LOV STACKING: submd should be from the subobj */
2197                 submd.lsm_object_id = loi->loi_id;
2198                 submd.lsm_stripe_count = 0;
2199                 lov_flags = *flags;
2200                 /* XXX submd is not fully initialized here */
2201                 rc = obd_match(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2202                                type, &sub_ext, mode, &lov_flags, data,
2203                                lov_lockhp);
2204                 if (rc != 1)
2205                         break;
2206         }
2207         if (rc == 1) {
2208                 if (lsm->lsm_stripe_count > 1) {
2209                         if (*flags & LDLM_FL_TEST_LOCK)
2210                                 lov_llh_destroy(lov_lockh);
2211                         lov_llh_put(lov_lockh);
2212                 }
2213                 RETURN(1);
2214         }
2215
2216         while (loi--, lov_lockhp--, i-- > 0) {
2217                 struct lov_stripe_md submd;
2218                 int err;
2219
2220                 if (lov_lockhp->cookie == 0)
2221                         continue;
2222
2223                 /* XXX LOV STACKING: submd should be from the subobj */
2224                 submd.lsm_object_id = loi->loi_id;
2225                 submd.lsm_stripe_count = 0;
2226                 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2227                                  mode, lov_lockhp);
2228                 if (err && lov->tgts[loi->loi_ost_idx].active) {
2229                         CERROR("error: cancelling objid "LPX64" on OST "
2230                                "idx %d after match failure: rc = %d\n",
2231                                loi->loi_id, loi->loi_ost_idx, err);
2232                 }
2233         }
2234
2235         if (lsm->lsm_stripe_count > 1) {
2236                 lov_llh_destroy(lov_lockh);
2237                 lov_llh_put(lov_lockh);
2238         }
2239         RETURN(rc);
2240 }
2241
2242 static int lov_change_cbdata(struct obd_export *exp,
2243                              struct lov_stripe_md *lsm, ldlm_iterator_t it,
2244                              void *data)
2245 {
2246         struct lov_obd *lov;
2247         struct lov_oinfo *loi;
2248         int rc = 0, i;
2249         ENTRY;
2250
2251         if (lsm_bad_magic(lsm))
2252                 RETURN(-EINVAL);
2253
2254         if (!exp || !exp->exp_obd)
2255                 RETURN(-ENODEV);
2256
2257         lov = &exp->exp_obd->u.lov;
2258         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2259                 struct lov_stripe_md submd;
2260                 if (lov->tgts[loi->loi_ost_idx].active == 0)
2261                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2262
2263                 submd.lsm_object_id = loi->loi_id;
2264                 submd.lsm_stripe_count = 0;
2265                 rc = obd_change_cbdata(lov->tgts[loi->loi_ost_idx].ltd_exp,
2266                                        &submd, it, data);
2267         }
2268         RETURN(rc);
2269 }
2270
2271 static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm,
2272                       __u32 mode, struct lustre_handle *lockh)
2273 {
2274         struct lov_lock_handles *lov_lockh = NULL;
2275         struct lustre_handle *lov_lockhp;
2276         struct lov_obd *lov;
2277         struct lov_oinfo *loi;
2278         int rc = 0, i;
2279         ENTRY;
2280
2281         if (lsm_bad_magic(lsm))
2282                 RETURN(-EINVAL);
2283
2284         if (!exp || !exp->exp_obd)
2285                 RETURN(-ENODEV);
2286
2287         LASSERT(lockh);
2288         if (lsm->lsm_stripe_count > 1) {
2289                 lov_lockh = lov_handle2llh(lockh);
2290                 if (!lov_lockh) {
2291                         CERROR("LOV: invalid lov lock handle %p\n", lockh);
2292                         RETURN(-EINVAL);
2293                 }
2294
2295                 lov_lockhp = lov_lockh->llh_handles;
2296         } else {
2297                 lov_lockhp = lockh;
2298         }
2299
2300         lov = &exp->exp_obd->u.lov;
2301         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2302              i++, loi++, lov_lockhp++) {
2303                 struct lov_stripe_md submd;
2304                 int err;
2305
2306                 if (lov_lockhp->cookie == 0) {
2307                         CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2308                                loi->loi_ost_idx, loi->loi_id);
2309                         continue;
2310                 }
2311
2312                 /* XXX LOV STACKING: submd should be from the subobj */
2313                 submd.lsm_object_id = loi->loi_id;
2314                 submd.lsm_stripe_count = 0;
2315                 err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd,
2316                                  mode, lov_lockhp);
2317                 if (err) {
2318                         if (lov->tgts[loi->loi_ost_idx].active) {
2319                                 CERROR("error: cancel objid "LPX64" subobj "
2320                                        LPX64" on OST idx %d: rc = %d\n",
2321                                        lsm->lsm_object_id,
2322                                        loi->loi_id, loi->loi_ost_idx, err);
2323                                 if (!rc)
2324                                         rc = err;
2325                         }
2326                 }
2327         }
2328
2329         if (lsm->lsm_stripe_count > 1)
2330                 lov_llh_destroy(lov_lockh);
2331         if (lov_lockh != NULL)
2332                 lov_llh_put(lov_lockh);
2333         RETURN(rc);
2334 }
2335
2336 static int lov_cancel_unused(struct obd_export *exp,
2337                              struct lov_stripe_md *lsm, int flags, void *opaque)
2338 {
2339         struct lov_obd *lov;
2340         struct lov_oinfo *loi;
2341         int rc = 0, i;
2342         ENTRY;
2343
2344         if (lsm_bad_magic(lsm))
2345                 RETURN(-EINVAL);
2346
2347         if (!exp || !exp->exp_obd)
2348                 RETURN(-ENODEV);
2349
2350         lov = &exp->exp_obd->u.lov;
2351         for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) {
2352                 struct lov_stripe_md submd;
2353                 int err;
2354
2355                 if (lov->tgts[loi->loi_ost_idx].active == 0)
2356                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2357
2358                 submd.lsm_object_id = loi->loi_id;
2359                 submd.lsm_stripe_count = 0;
2360                 err = obd_cancel_unused(lov->tgts[loi->loi_ost_idx].ltd_exp,
2361                                         &submd, flags, opaque);
2362                 if (err && lov->tgts[loi->loi_ost_idx].active) {
2363                         CERROR("error: cancel unused objid "LPX64" subobj "LPX64
2364                                " on OST idx %d: rc = %d\n", lsm->lsm_object_id,
2365                                loi->loi_id, loi->loi_ost_idx, err);
2366                         if (!rc)
2367                                 rc = err;
2368                 }
2369         }
2370         RETURN(rc);
2371 }
2372
2373 #define LOV_U64_MAX ((__u64)~0ULL)
2374 #define LOV_SUM_MAX(tot, add)                                           \
2375         do {                                                            \
2376                 if ((tot) + (add) < (tot))                              \
2377                         (tot) = LOV_U64_MAX;                            \
2378                 else                                                    \
2379                         (tot) += (add);                                 \
2380         } while(0)
2381
2382 static int lov_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2383                       unsigned long max_age)
2384 {
2385         struct lov_obd *lov = &obd->u.lov;
2386         struct obd_statfs lov_sfs;
2387         int set = 0;
2388         int rc = 0;
2389         int i;
2390         ENTRY;
2391
2392
2393         /* We only get block data from the OBD */
2394         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2395                 int err;
2396
2397                 if (!lov->tgts[i].active) {
2398                         CDEBUG(D_HA, "lov idx %d inactive\n", i);
2399                         continue;
2400                 }
2401
2402                 err = obd_statfs(class_exp2obd(lov->tgts[i].ltd_exp), &lov_sfs,
2403                                  max_age);
2404                 if (err) {
2405                         if (lov->tgts[i].active && !rc)
2406                                 rc = err;
2407                         continue;
2408                 }
2409
2410                 if (!set) {
2411                         memcpy(osfs, &lov_sfs, sizeof(lov_sfs));
2412                         set = 1;
2413                 } else {
2414                         osfs->os_bfree += lov_sfs.os_bfree;
2415                         osfs->os_bavail += lov_sfs.os_bavail;
2416                         osfs->os_blocks += lov_sfs.os_blocks;
2417                         /* XXX not sure about this one - depends on policy.
2418                          *   - could be minimum if we always stripe on all OBDs
2419                          *     (but that would be wrong for any other policy,
2420                          *     if one of the OBDs has no more objects left)
2421                          *   - could be sum if we stripe whole objects
2422                          *   - could be average, just to give a nice number
2423                          *
2424                          * To give a "reasonable" (if not wholly accurate)
2425                          * number, we divide the total number of free objects
2426                          * by expected stripe count (watch out for overflow).
2427                          */
2428                         LOV_SUM_MAX(osfs->os_files, lov_sfs.os_files);
2429                         LOV_SUM_MAX(osfs->os_ffree, lov_sfs.os_ffree);
2430                 }
2431         }
2432
2433         if (set) {
2434                 __u32 expected_stripes = lov->desc.ld_default_stripe_count ?
2435                                          lov->desc.ld_default_stripe_count :
2436                                          lov->desc.ld_active_tgt_count;
2437
2438                 if (osfs->os_files != LOV_U64_MAX)
2439                         do_div(osfs->os_files, expected_stripes);
2440                 if (osfs->os_ffree != LOV_U64_MAX)
2441                         do_div(osfs->os_ffree, expected_stripes);
2442         } else if (!rc)
2443                 rc = -EIO;
2444
2445         RETURN(rc);
2446 }
2447
2448 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2449                          void *karg, void *uarg)
2450 {
2451         struct obd_device *obddev = class_exp2obd(exp);
2452         struct lov_obd *lov = &obddev->u.lov;
2453         int i, count = lov->desc.ld_tgt_count;
2454         struct obd_uuid *uuidp;
2455         int rc;
2456
2457         ENTRY;
2458
2459         switch (cmd) {
2460         case OBD_IOC_LOV_GET_CONFIG: {
2461                 struct obd_ioctl_data *data = karg;
2462                 struct lov_tgt_desc *tgtdesc;
2463                 struct lov_desc *desc;
2464                 char *buf = NULL;
2465
2466                 buf = NULL;
2467                 len = 0;
2468                 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2469                         RETURN(-EINVAL);
2470
2471                 data = (struct obd_ioctl_data *)buf;
2472
2473                 if (sizeof(*desc) > data->ioc_inllen1) {
2474                         OBD_FREE(buf, len);
2475                         RETURN(-EINVAL);
2476                 }
2477
2478                 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
2479                         OBD_FREE(buf, len);
2480                         RETURN(-EINVAL);
2481                 }
2482
2483                 desc = (struct lov_desc *)data->ioc_inlbuf1;
2484                 memcpy(desc, &(lov->desc), sizeof(*desc));
2485
2486                 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
2487                 tgtdesc = lov->tgts;
2488                 for (i = 0; i < count; i++, uuidp++, tgtdesc++)
2489                         obd_str2uuid(uuidp, tgtdesc->uuid.uuid);
2490
2491                 rc = copy_to_user((void *)uarg, buf, len);
2492                 if (rc)
2493                         rc = -EFAULT;
2494                 obd_ioctl_freedata(buf, len);
2495                 break;
2496         }
2497         case LL_IOC_LOV_SETSTRIPE:
2498                 rc = lov_setstripe(exp, karg, uarg);
2499                 break;
2500         case LL_IOC_LOV_GETSTRIPE:
2501                 rc = lov_getstripe(exp, karg, uarg);
2502                 break;
2503         case LL_IOC_LOV_SETEA:
2504                 rc = lov_setea(exp, karg, uarg);
2505                 break;
2506         default: {
2507                 int set = 0;
2508                 if (count == 0)
2509                         RETURN(-ENOTTY);
2510                 rc = 0;
2511                 for (i = 0; i < count; i++) {
2512                         int err;
2513
2514                         err = obd_iocontrol(cmd, lov->tgts[i].ltd_exp,
2515                                             len, karg, uarg);
2516                         if (err) {
2517                                 if (lov->tgts[i].active) {
2518                                         CERROR("error: iocontrol OSC %s on OST"
2519                                                "idx %d: err = %d\n",
2520                                                lov->tgts[i].uuid.uuid, i, err);
2521                                         if (!rc)
2522                                                 rc = err;
2523                                 }
2524                         } else
2525                                 set = 1;
2526                 }
2527                 if (!set && !rc)
2528                         rc = -EIO;
2529         }
2530         }
2531
2532         RETURN(rc);
2533 }
2534
2535 static int lov_get_info(struct obd_export *exp, __u32 keylen,
2536                         void *key, __u32 *vallen, void *val)
2537 {
2538         struct obd_device *obddev = class_exp2obd(exp);
2539         struct lov_obd *lov = &obddev->u.lov;
2540         int i;
2541         ENTRY;
2542
2543         if (!vallen || !val)
2544                 RETURN(-EFAULT);
2545
2546         if (keylen > strlen("lock_to_stripe") &&
2547             strcmp(key, "lock_to_stripe") == 0) {
2548                 struct {
2549                         char name[16];
2550                         struct ldlm_lock *lock;
2551                         struct lov_stripe_md *lsm;
2552                 } *data = key;
2553                 struct lov_oinfo *loi;
2554                 __u32 *stripe = val;
2555
2556                 if (*vallen < sizeof(*stripe))
2557                         RETURN(-EFAULT);
2558                 *vallen = sizeof(*stripe);
2559
2560                 /* XXX This is another one of those bits that will need to
2561                  * change if we ever actually support nested LOVs.  It uses
2562                  * the lock's export to find out which stripe it is. */
2563                 for (i = 0, loi = data->lsm->lsm_oinfo;
2564                      i < data->lsm->lsm_stripe_count;
2565                      i++, loi++) {
2566                         if (lov->tgts[loi->loi_ost_idx].ltd_exp ==
2567                             data->lock->l_conn_export) {
2568                                 *stripe = i;
2569                                 RETURN(0);
2570                         }
2571                 }
2572                 RETURN(-ENXIO);
2573         } else if (keylen >= strlen("size_to_stripe") &&
2574                    strcmp(key, "size_to_stripe") == 0) {
2575                 struct {
2576                         int stripe_number;
2577                         __u64 size;
2578                         struct lov_stripe_md *lsm;
2579                 } *data = val;
2580
2581                 if (*vallen < sizeof(*data))
2582                         RETURN(-EFAULT);
2583
2584                 data->size = lov_size_to_stripe(data->lsm, data->size,
2585                                                 data->stripe_number);
2586                 RETURN(0);
2587         } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2588                 obd_id *ids = val;
2589                 int rc, size = sizeof(obd_id);
2590                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2591                         if (!lov->tgts[i].active)
2592                                 continue;
2593                         rc = obd_get_info(lov->tgts[i].ltd_exp, keylen, key,
2594                                           &size, &(ids[i]));
2595                         if (rc != 0)
2596                                 RETURN(rc);
2597                 }
2598                 RETURN(0);
2599         } else if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
2600                 struct lov_desc *desc_ret = val;
2601                 *desc_ret = lov->desc;
2602
2603                 RETURN(0);
2604         }
2605
2606         RETURN(-EINVAL);
2607 }
2608
2609 static int lov_set_info(struct obd_export *exp, obd_count keylen,
2610                         void *key, obd_count vallen, void *val)
2611 {
2612         struct obd_device *obddev = class_exp2obd(exp);
2613         struct lov_obd *lov = &obddev->u.lov;
2614         int i, rc = 0;
2615         ENTRY;
2616
2617 #define KEY_IS(str) \
2618         (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
2619
2620         if (KEY_IS("next_id")) {
2621                 if (vallen != lov->desc.ld_tgt_count)
2622                         RETURN(-EINVAL);
2623                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2624                         int er;
2625
2626                         /* initialize all OSCs, even inactive ones */
2627
2628                         er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key,
2629                                           sizeof(obd_id), ((obd_id*)val) + i);
2630                         if (!rc)
2631                                 rc = er;
2632                 }
2633                 RETURN(rc);
2634         }
2635
2636         if (KEY_IS("growth_count")) {
2637                 if (vallen != sizeof(int))
2638                         RETURN(-EINVAL);
2639         } else if (KEY_IS("mds_conn") || KEY_IS("unlinked")) {
2640                 if (vallen != 0)
2641                         RETURN(-EINVAL);
2642         } else {
2643                 RETURN(-EINVAL);
2644         }
2645
2646         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2647                 int er;
2648
2649                 if (val && !obd_uuid_equals(val, &lov->tgts[i].uuid))
2650                         continue;
2651
2652                 if (!val && !lov->tgts[i].active)
2653                         continue;
2654
2655                 er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key, vallen,
2656                                    val);
2657                 if (!rc)
2658                         rc = er;
2659         }
2660         RETURN(rc);
2661 #undef KEY_IS
2662
2663 }
2664
2665 /* Merge rss if kms == 0
2666  *
2667  * Even when merging RSS, we will take the KMS value if it's larger.
2668  * This prevents getattr from stomping on dirty cached pages which
2669  * extend the file size. */
2670 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms)
2671 {
2672         struct lov_oinfo *loi;
2673         __u64 size = 0;
2674         int i;
2675
2676         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2677              i++, loi++) {
2678                 obd_size lov_size, tmpsize;
2679
2680                 tmpsize = loi->loi_kms;
2681                 if (kms == 0 && loi->loi_rss > tmpsize)
2682                         tmpsize = loi->loi_rss;
2683
2684                 lov_size = lov_stripe_size(lsm, tmpsize, i);
2685                 if (lov_size > size)
2686                         size = lov_size;
2687         }
2688         return size;
2689 }
2690 EXPORT_SYMBOL(lov_merge_size);
2691
2692 __u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time)
2693 {
2694         struct lov_oinfo *loi;
2695         int i;
2696
2697         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2698              i++, loi++) {
2699                 if (loi->loi_mtime > current_time)
2700                         current_time = loi->loi_mtime;
2701         }
2702         return current_time;
2703 }
2704 EXPORT_SYMBOL(lov_merge_mtime);
2705
2706 #if 0
2707 struct lov_multi_wait {
2708         struct ldlm_lock *lock;
2709         wait_queue_t      wait;
2710         int               completed;
2711         int               generation;
2712 };
2713
2714 int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm,
2715                       struct lustre_handle *lockh)
2716 {
2717         struct lov_lock_handles *lov_lockh = NULL;
2718         struct lustre_handle *lov_lockhp;
2719         struct lov_obd *lov;
2720         struct lov_oinfo *loi;
2721         struct lov_multi_wait *queues;
2722         int rc = 0, i;
2723         ENTRY;
2724
2725         if (lsm_bad_magic(lsm))
2726                 RETURN(-EINVAL);
2727
2728         if (!exp || !exp->exp_obd)
2729                 RETURN(-ENODEV);
2730
2731         LASSERT(lockh != NULL);
2732         if (lsm->lsm_stripe_count > 1) {
2733                 lov_lockh = lov_handle2llh(lockh);
2734                 if (lov_lockh == NULL) {
2735                         CERROR("LOV: invalid lov lock handle %p\n", lockh);
2736                         RETURN(-EINVAL);
2737                 }
2738
2739                 lov_lockhp = lov_lockh->llh_handles;
2740         } else {
2741                 lov_lockhp = lockh;
2742         }
2743
2744         OBD_ALLOC(queues, lsm->lsm_stripe_count * sizeof(*queues));
2745         if (queues == NULL)
2746                 GOTO(out, rc = -ENOMEM);
2747
2748         lov = &exp->exp_obd->u.lov;
2749         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2750              i++, loi++, lov_lockhp++) {
2751                 struct ldlm_lock *lock;
2752                 struct obd_device *obd;
2753                 unsigned long irqflags;
2754
2755                 lock = ldlm_handle2lock(lov_lockhp);
2756                 if (lock == NULL) {
2757                         CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2758                                loi->loi_ost_idx, loi->loi_id);
2759                         queues[i].completed = 1;
2760                         continue;
2761                 }
2762
2763                 queues[i].lock = lock;
2764                 init_waitqueue_entry(&(queues[i].wait), current);
2765                 add_wait_queue(lock->l_waitq, &(queues[i].wait));
2766
2767                 obd = class_exp2obd(lock->l_conn_export);
2768                 if (obd != NULL)
2769                         imp = obd->u.cli.cl_import;
2770                 if (imp != NULL) {
2771                         spin_lock_irqsave(&imp->imp_lock, irqflags);
2772                         queues[i].generation = imp->imp_generation;
2773                         spin_unlock_irqrestore(&imp->imp_lock, irqflags);
2774                 }
2775         }
2776
2777         lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
2778                                interrupted_completion_wait, &lwd);
2779         rc = l_wait_event_added(check_multi_complete(queues, lsm), &lwi);
2780
2781         for (i = 0; i < lsm->lsm_stripe_count; i++)
2782                 remove_wait_queue(lock->l_waitq, &(queues[i].wait));
2783
2784         if (rc == -EINTR || rc == -ETIMEDOUT) {
2785
2786
2787         }
2788
2789  out:
2790         if (lov_lockh != NULL)
2791                 lov_llh_put(lov_lockh);
2792         RETURN(rc);
2793 }
2794 #endif
2795
2796 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
2797                       obd_off size)
2798 {
2799         struct lov_oinfo *loi;
2800         int stripe = 0;
2801         __u64 kms;
2802         ENTRY;
2803
2804         if (size > 0)
2805                 stripe = lov_stripe_number(lsm, size - 1);
2806         kms = lov_size_to_stripe(lsm, size, stripe);
2807         loi = &(lsm->lsm_oinfo[stripe]);
2808
2809         CDEBUG(D_INODE, "stripe %d KMS %sincreasing "LPU64"->"LPU64"\n",
2810                stripe, kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms);
2811         if (kms > loi->loi_kms)
2812                 loi->loi_kms = kms;
2813         EXIT;
2814 }
2815 EXPORT_SYMBOL(lov_increase_kms);
2816
2817 struct obd_ops lov_obd_ops = {
2818         o_owner:       THIS_MODULE,
2819         o_attach:      lov_attach,
2820         o_detach:      lov_detach,
2821         o_setup:       lov_setup,
2822         o_cleanup:     lov_cleanup,
2823         o_connect:     lov_connect,
2824         o_disconnect:  lov_disconnect,
2825         o_statfs:      lov_statfs,
2826         o_packmd:      lov_packmd,
2827         o_unpackmd:    lov_unpackmd,
2828         o_create:      lov_create,
2829         o_destroy:     lov_destroy,
2830         o_getattr:     lov_getattr,
2831         o_getattr_async: lov_getattr_async,
2832         o_setattr:     lov_setattr,
2833         o_brw:         lov_brw,
2834         o_brw_async:   lov_brw_async,
2835         .o_prep_async_page =    lov_prep_async_page,
2836         .o_queue_async_io =     lov_queue_async_io,
2837         .o_set_async_flags =    lov_set_async_flags,
2838         .o_queue_group_io =     lov_queue_group_io,
2839         .o_trigger_group_io =   lov_trigger_group_io,
2840         .o_teardown_async_page  lov_teardown_async_page,
2841         o_punch:       lov_punch,
2842         o_sync:        lov_sync,
2843         o_enqueue:     lov_enqueue,
2844         o_match:       lov_match,
2845         o_change_cbdata: lov_change_cbdata,
2846         o_cancel:      lov_cancel,
2847         o_cancel_unused: lov_cancel_unused,
2848         o_iocontrol:   lov_iocontrol,
2849         o_get_info:    lov_get_info,
2850         o_set_info:    lov_set_info,
2851         o_llog_init:   lov_llog_init,
2852         o_llog_finish: lov_llog_finish,
2853         o_notify: lov_notify,
2854 };
2855
2856 int __init lov_init(void)
2857 {
2858         struct lprocfs_static_vars lvars;
2859         int rc;
2860
2861         lprocfs_init_vars(lov, &lvars);
2862         rc = class_register_type(&lov_obd_ops, lvars.module_vars,
2863                                  OBD_LOV_DEVICENAME);
2864         RETURN(rc);
2865 }
2866
2867 #ifdef __KERNEL__
2868 static void /*__exit*/ lov_exit(void)
2869 {
2870         class_unregister_type(OBD_LOV_DEVICENAME);
2871 }
2872
2873 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2874 MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver");
2875 MODULE_LICENSE("GPL");
2876
2877 module_init(lov_init);
2878 module_exit(lov_exit);
2879 #endif