lustre/osc/osc_create.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/osc/osc_create.c
  37  * For testing and management it is treated as an obd_device,
  38  * although * it does not export a full OBD method table (the
  39  * requests are coming * in over the wire, so object target modules
  40  * do not have a full * method table.)
  41  *
  42  * Author: Peter Braam <braam@clusterfs.com>
  43  */
  44
  45 #ifndef EXPORT_SYMTAB
  46 # define EXPORT_SYMTAB
  47 #endif
  48 #define DEBUG_SUBSYSTEM S_OSC
  49
  50 #ifdef __KERNEL__
  51 # include <libcfs/libcfs.h>
  52 #else /* __KERNEL__ */
  53 # include <liblustre.h>
  54 #endif
  55
  56 #ifdef  __CYGWIN__
  57 # include <ctype.h>
  58 #endif
  59
  60 #include <lustre_dlm.h>
  61 #include <obd_class.h>
  62 #include "osc_internal.h"
  63
  64 /* XXX need AT adjust ? */
  65 #define osc_create_timeout      (obd_timeout / 2)
  66
  67 struct osc_create_async_args {
  68         struct osc_creator      *rq_oscc;
  69         struct lov_stripe_md    *rq_lsm;
  70         struct obd_info         *rq_oinfo;
  71 };
  72
  73 static int oscc_internal_create(struct osc_creator *oscc);
  74 static int handle_async_create(struct ptlrpc_request *req, int rc);
  75
  76 static int osc_interpret_create(const struct lu_env *env,
  77                                 struct ptlrpc_request *req, void *data, int rc)
  78 {
  79         struct osc_creator *oscc;
  80         struct ost_body *body = NULL;
  81         struct ptlrpc_request *fake_req, *pos;
  82         ENTRY;
  83
  84         if (req->rq_repmsg) {
  85                 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
  86                 if (body == NULL && rc == 0)
  87                         rc = -EPROTO;
  88         }
  89
  90         oscc = req->rq_async_args.pointer_arg[0];
  91         LASSERT(oscc && (oscc->oscc_obd != LP_POISON));
  92
  93         cfs_spin_lock(&oscc->oscc_lock);
  94         oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
  95         switch (rc) {
  96         case 0: {
  97                 if (body) {
  98                         int diff =ostid_id(&body->oa.o_oi)- oscc->oscc_last_id;
  99
 100                         /* oscc_internal_create() stores the original value of
 101                          * grow_count in rq_async_args.space[0].
 102                          * We can't compare against oscc_grow_count directly,
 103                          * because it may have been increased while the RPC
 104                          * is in flight, so we would always find ourselves
 105                          * having created fewer objects and decreasing the
 106                          * precreate request size.  b=18577 */
 107                         if (diff < (int) req->rq_async_args.space[0]) {
 108                                 /* the OST has not managed to create all the
 109                                  * objects we asked for */
 110                                 oscc->oscc_grow_count = max(diff,
 111                                                             OST_MIN_PRECREATE);
 112                                 /* don't bump grow_count next time */
 113                                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 114                         } else {
 115                                 /* the OST is able to keep up with the work,
 116                                  * we could consider increasing grow_count
 117                                  * next time if needed */
 118                                 oscc->oscc_flags &= ~OSCC_FLAG_LOW;
 119                         }
 120                         oscc->oscc_last_id = ostid_id(&body->oa.o_oi);
 121                 }
 122                 cfs_spin_unlock(&oscc->oscc_lock);
 123                 break;
 124         }
 125         case -EROFS:
 126                 oscc->oscc_flags |= OSCC_FLAG_RDONLY;
 127         case -ENOSPC:
 128         case -EFBIG:
 129                 if (rc != -EROFS) {
 130                         oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 131                         if (body && rc == -ENOSPC) {
 132                                 oscc->oscc_last_id = body->oa.o_id;
 133                                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
 134                         }
 135                 }
 136                 cfs_spin_unlock(&oscc->oscc_lock);
 137                 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
 138                 break;
 139         case -EIO: {
 140                 /* filter always set body->oa.o_id as the last_id
 141                  * of filter (see filter_handle_precreate for detail)*/
 142                 if (body && body->oa.o_id > oscc->oscc_last_id)
 143                         oscc->oscc_last_id = body->oa.o_id;
 144                 cfs_spin_unlock(&oscc->oscc_lock);
 145                 break;
 146         }
 147         case -EINTR:
 148         case -EWOULDBLOCK: {
 149                 /* aka EAGAIN we should not delay create if import failed -
 150                  * this avoid client stick in create and avoid race with
 151                  * delorphan */
 152                 /* EINTR say - old create request is killed due mds<>ost
 153                  * eviction - OSCC_FLAG_RECOVERING can already set due
 154                  * IMP_DISCONN event */
 155                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 156                 /* oscc->oscc_grow_count = OST_MIN_PRECREATE; */
 157                 cfs_spin_unlock(&oscc->oscc_lock);
 158                 break;
 159         }
 160         default: {
 161                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 162                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
 163                 cfs_spin_unlock(&oscc->oscc_lock);
 164                 DEBUG_REQ(D_ERROR, req,
 165                           "Unknown rc %d from async create: failing oscc", rc);
 166                 ptlrpc_fail_import(req->rq_import,
 167                                    lustre_msg_get_conn_cnt(req->rq_reqmsg));
 168         }
 169         }
 170
 171         CDEBUG(D_HA, "preallocated through id "LPU64" (next to use "LPU64")\n",
 172                oscc->oscc_last_id, oscc->oscc_next_id);
 173
 174         cfs_spin_lock(&oscc->oscc_lock);
 175         cfs_list_for_each_entry_safe(fake_req, pos,
 176                                      &oscc->oscc_wait_create_list, rq_list) {
 177                 if (handle_async_create(fake_req, rc)  == -EAGAIN) {
 178                         oscc_internal_create(oscc);
 179                         /* sending request should be never fail because
 180                          * osc use preallocated requests pool */
 181                         GOTO(exit_wakeup, rc);
 182                 }
 183         }
 184         cfs_spin_unlock(&oscc->oscc_lock);
 185
 186 exit_wakeup:
 187         cfs_waitq_signal(&oscc->oscc_waitq);
 188         RETURN(rc);
 189 }
 190
 191 static int oscc_internal_create(struct osc_creator *oscc)
 192 {
 193         struct ptlrpc_request *request;
 194         struct ost_body *body;
 195         ENTRY;
 196
 197         LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
 198
 199         /* Do not check for a degraded OST here - bug21563/bug18539 */
 200         if (oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 201                 cfs_spin_unlock(&oscc->oscc_lock);
 202                 RETURN(0);
 203         }
 204
 205         /* we need check it before OSCC_FLAG_CREATING - because need
 206          * see lower number of precreate objects */
 207         if (oscc->oscc_grow_count < oscc->oscc_max_grow_count &&
 208             ((oscc->oscc_flags & OSCC_FLAG_LOW) == 0) &&
 209             (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
 210                    (oscc->oscc_grow_count / 4 + 1)) {
 211                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 212                 oscc->oscc_grow_count *= 2;
 213         }
 214
 215         if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
 216                 cfs_spin_unlock(&oscc->oscc_lock);
 217                 RETURN(0);
 218         }
 219
 220         if (oscc->oscc_grow_count > oscc->oscc_max_grow_count / 2)
 221                 oscc->oscc_grow_count = oscc->oscc_max_grow_count / 2;
 222
 223         oscc->oscc_flags |= OSCC_FLAG_CREATING;
 224         cfs_spin_unlock(&oscc->oscc_lock);
 225
 226         request = ptlrpc_request_alloc_pack(oscc->oscc_obd->u.cli.cl_import,
 227                                             &RQF_OST_CREATE,
 228                                             LUSTRE_OST_VERSION, OST_CREATE);
 229         if (request == NULL) {
 230                 cfs_spin_lock(&oscc->oscc_lock);
 231                 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
 232                 cfs_spin_unlock(&oscc->oscc_lock);
 233                 RETURN(-ENOMEM);
 234         }
 235
 236         request->rq_request_portal = OST_CREATE_PORTAL;
 237         ptlrpc_at_set_req_timeout(request);
 238         body = req_capsule_client_get(&request->rq_pill, &RMF_OST_BODY);
 239
 240         cfs_spin_lock(&oscc->oscc_lock);
 241
 242         if (likely(fid_seq_is_mdt(oscc->oscc_oa.o_seq))) {
 243                 body->oa.o_oi.oi_seq = oscc->oscc_oa.o_seq;
 244                 body->oa.o_oi.oi_id  = oscc->oscc_last_id +
 245                                        oscc->oscc_grow_count;
 246         } else {
 247                 /*Just warning here currently, since not sure how fid-on-ost
 248                  *will be implemented here */
 249                 CWARN("o_seq: "LPU64" is not indicate any MDTs.\n",
 250                        oscc->oscc_oa.o_seq);
 251         }
 252
 253         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 254         request->rq_async_args.space[0] = oscc->oscc_grow_count;
 255         cfs_spin_unlock(&oscc->oscc_lock);
 256         CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
 257                body->oa.o_id, oscc->oscc_last_id);
 258
 259         /* we should not resend create request - anyway we will have delorphan
 260          * and kill these objects */
 261         request->rq_no_delay = request->rq_no_resend = 1;
 262         ptlrpc_request_set_replen(request);
 263
 264         request->rq_async_args.pointer_arg[0] = oscc;
 265         request->rq_interpret_reply = osc_interpret_create;
 266         ptlrpcd_add_req(request, PSCOPE_OTHER);
 267
 268         RETURN(0);
 269 }
 270
 271 static int oscc_has_objects_nolock(struct osc_creator *oscc, int count)
 272 {
 273         return ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 274 }
 275
 276
 277 static int oscc_has_objects(struct osc_creator *oscc, int count)
 278 {
 279         int have_objs;
 280
 281         cfs_spin_lock(&oscc->oscc_lock);
 282         have_objs = oscc_has_objects_nolock(oscc, count);
 283         cfs_spin_unlock(&oscc->oscc_lock);
 284
 285         return have_objs;
 286 }
 287
 288 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 289 {
 290         int have_objs;
 291         int ost_unusable;
 292
 293         ost_unusable = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 294
 295         cfs_spin_lock(&oscc->oscc_lock);
 296         ost_unusable |= (OSCC_FLAG_NOSPC | OSCC_FLAG_RDONLY |
 297                          OSCC_FLAG_EXITING) & oscc->oscc_flags;
 298         have_objs = oscc_has_objects_nolock(oscc, count);
 299
 300         if (!ost_unusable && !have_objs)
 301                 /* they release lock himself */
 302                 have_objs = oscc_internal_create(oscc);
 303         else
 304                 cfs_spin_unlock(&oscc->oscc_lock);
 305
 306         return have_objs || ost_unusable;
 307 }
 308
 309 static int oscc_precreate(struct osc_creator *oscc)
 310 {
 311         struct l_wait_info lwi;
 312         int rc = 0;
 313         ENTRY;
 314
 315         if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
 316                 RETURN(0);
 317
 318         /* we should be not block forever - because client's create rpc can
 319          * stick in mds for long time and forbid client reconnect */
 320         lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(osc_create_timeout)),
 321                           NULL, NULL);
 322
 323         rc = l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 324         RETURN(rc);
 325 }
 326
 327 static int oscc_in_sync(struct osc_creator *oscc)
 328 {
 329         int sync;
 330
 331         cfs_spin_lock(&oscc->oscc_lock);
 332         sync = oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS;
 333         cfs_spin_unlock(&oscc->oscc_lock);
 334
 335         return sync;
 336 }
 337
 338 /* decide if the OST has remaining object, return value :
 339         0 : the OST has remaining object, and don't need to do precreate.
 340         1 : the OST has no remaining object, and will send a RPC for precreate.
 341         2 : the OST has no remaining object, and will not get any for
 342             a potentially very long time
 343      1000 : unusable
 344  */
 345 int osc_precreate(struct obd_export *exp)
 346 {
 347         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 348         struct obd_import *imp = exp->exp_imp_reverse;
 349         ENTRY;
 350
 351         LASSERT(oscc != NULL);
 352         if (imp != NULL && imp->imp_deactive)
 353                 RETURN(1000);
 354
 355         /* Handle critical states first */
 356         cfs_spin_lock(&oscc->oscc_lock);
 357         if (oscc->oscc_flags & OSCC_FLAG_NOSPC ||
 358             oscc->oscc_flags & OSCC_FLAG_RDONLY ||
 359             oscc->oscc_flags & OSCC_FLAG_EXITING) {
 360                 cfs_spin_unlock(&oscc->oscc_lock);
 361                 RETURN(1000);
 362         }
 363
 364         if ((oscc->oscc_flags & OSCC_FLAG_RECOVERING) ||
 365             (oscc->oscc_flags & OSCC_FLAG_DEGRADED)) {
 366                 cfs_spin_unlock(&oscc->oscc_lock);
 367                 RETURN(2);
 368         }
 369
 370         if (oscc_has_objects_nolock(oscc, oscc->oscc_grow_count / 2)) {
 371                 cfs_spin_unlock(&oscc->oscc_lock);
 372                 RETURN(0);
 373         }
 374
 375         /* Do not check for OSCC_FLAG_CREATING flag here, let
 376          * osc_precreate() call oscc_internal_create() and
 377          * adjust oscc_grow_count bug21563 */
 378         if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 379                 cfs_spin_unlock(&oscc->oscc_lock);
 380                 RETURN(1);
 381         }
 382
 383         if (oscc_internal_create(oscc))
 384                 RETURN(1000);
 385         RETURN(1);
 386 }
 387
 388 static int handle_async_create(struct ptlrpc_request *req, int rc)
 389 {
 390         struct osc_create_async_args *args = ptlrpc_req_async_args(req);
 391         struct osc_creator    *oscc = args->rq_oscc;
 392         struct lov_stripe_md  *lsm  = args->rq_lsm;
 393         struct obd_info       *oinfo = args->rq_oinfo;
 394         struct obdo           *oa = oinfo->oi_oa;
 395
 396         LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
 397
 398         if(rc)
 399                 GOTO(out_wake, rc);
 400
 401         /* Handle the critical type errors first.
 402          * Should we also test cl_import state as well ? */
 403         if (oscc->oscc_flags & OSCC_FLAG_EXITING)
 404                 GOTO(out_wake, rc = -EIO);
 405
 406         if (oscc->oscc_flags & OSCC_FLAG_NOSPC)
 407                 GOTO(out_wake, rc = -ENOSPC);
 408
 409         if (oscc->oscc_flags & OSCC_FLAG_RDONLY)
 410                 GOTO(out_wake, rc = -EROFS);
 411
 412         /* should be try wait until recovery finished */
 413         if((oscc->oscc_flags & OSCC_FLAG_RECOVERING) ||
 414            (oscc->oscc_flags & OSCC_FLAG_DEGRADED))
 415                 RETURN(-EAGAIN);
 416
 417         if (oscc_has_objects_nolock(oscc, 1)) {
 418                 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 419                 oa->o_id = oscc->oscc_next_id;
 420                 lsm->lsm_object_id = oscc->oscc_next_id;
 421                 oscc->oscc_next_id++;
 422
 423                 CDEBUG(D_RPCTRACE, " set oscc_next_id = "LPU64"\n",
 424                        oscc->oscc_next_id);
 425                 GOTO(out_wake, rc = 0);
 426         }
 427
 428         /* we don't have objects now - continue wait */
 429         RETURN(-EAGAIN);
 430
 431 out_wake:
 432
 433         rc = oinfo->oi_cb_up(oinfo, rc);
 434         ptlrpc_fakereq_finished(req);
 435
 436         RETURN(rc);
 437 }
 438
 439 static int async_create_interpret(const struct lu_env *env,
 440                                   struct ptlrpc_request *req, void *data,
 441                                   int rc)
 442 {
 443         struct osc_create_async_args *args = ptlrpc_req_async_args(req);
 444         struct osc_creator    *oscc = args->rq_oscc;
 445         int ret;
 446
 447         cfs_spin_lock(&oscc->oscc_lock);
 448         ret = handle_async_create(req, rc);
 449         cfs_spin_unlock(&oscc->oscc_lock);
 450
 451         return ret;
 452 }
 453
 454 int osc_create_async(struct obd_export *exp, struct obd_info *oinfo,
 455                      struct lov_stripe_md **ea, struct obd_trans_info *oti)
 456 {
 457         int rc;
 458         struct ptlrpc_request *fake_req;
 459         struct osc_create_async_args *args;
 460         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 461         struct obdo *oa = oinfo->oi_oa;
 462         ENTRY;
 463
 464         if ((oa->o_valid & OBD_MD_FLGROUP) && !fid_seq_is_mdt(oa->o_seq)) {
 465                 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
 466                 rc = oinfo->oi_cb_up(oinfo, rc);
 467                 RETURN(rc);
 468         }
 469
 470         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 471             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 472                 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
 473                 rc = oinfo->oi_cb_up(oinfo, rc);
 474                 RETURN(rc);
 475         }
 476
 477         LASSERT((*ea) != NULL);
 478
 479         fake_req = ptlrpc_prep_fakereq(oscc->oscc_obd->u.cli.cl_import,
 480                                        osc_create_timeout,
 481                                        async_create_interpret);
 482         if (fake_req == NULL) {
 483                 rc = oinfo->oi_cb_up(oinfo, -ENOMEM);
 484                 RETURN(-ENOMEM);
 485         }
 486
 487         args = ptlrpc_req_async_args(fake_req);
 488         CLASSERT(sizeof(*args) <= sizeof(fake_req->rq_async_args));
 489
 490         args->rq_oscc  = oscc;
 491         args->rq_lsm   = *ea;
 492         args->rq_oinfo = oinfo;
 493
 494         cfs_spin_lock(&oscc->oscc_lock);
 495         /* try fast path */
 496         rc = handle_async_create(fake_req, 0);
 497         if (rc == -EAGAIN) {
 498                 int is_add;
 499                 /* we not have objects - try wait */
 500                 is_add = ptlrpcd_add_req(fake_req, PSCOPE_OTHER);
 501                 if (!is_add)
 502                         cfs_list_add(&fake_req->rq_list,
 503                                      &oscc->oscc_wait_create_list);
 504                 else
 505                         rc = is_add;
 506         }
 507         cfs_spin_unlock(&oscc->oscc_lock);
 508
 509         if (rc != -EAGAIN)
 510                 /* need free request if was error hit or
 511                  * objects already allocated */
 512                 ptlrpc_req_finished(fake_req);
 513         else
 514                 /* EAGAIN mean - request is delayed */
 515                 rc = 0;
 516
 517         RETURN(rc);
 518 }
 519
 520 int osc_create(struct obd_export *exp, struct obdo *oa,
 521                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 522 {
 523         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 524         struct obd_import  *imp  = exp->exp_obd->u.cli.cl_import;
 525         struct lov_stripe_md *lsm;
 526         int del_orphan = 0, rc = 0;
 527         ENTRY;
 528
 529         LASSERT(oa);
 530         LASSERT(ea);
 531         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 532
 533         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 534             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 535                 RETURN(osc_real_create(exp, oa, ea, oti));
 536         }
 537
 538         if (!fid_seq_is_mdt(oa->o_seq))
 539                 RETURN(osc_real_create(exp, oa, ea, oti));
 540
 541         /* this is the special case where create removes orphans */
 542         if (oa->o_valid & OBD_MD_FLFLAGS &&
 543             oa->o_flags == OBD_FL_DELORPHAN) {
 544                 cfs_spin_lock(&oscc->oscc_lock);
 545                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 546                         cfs_spin_unlock(&oscc->oscc_lock);
 547                         RETURN(-EBUSY);
 548                 }
 549                 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 550                         cfs_spin_unlock(&oscc->oscc_lock);
 551                         RETURN(0);
 552                 }
 553
 554                 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
 555                 /* seting flag LOW we prevent extra grow precreate size
 556                  * and enforce use last assigned size */
 557                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 558                 cfs_spin_unlock(&oscc->oscc_lock);
 559                 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
 560                        oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
 561
 562                 del_orphan = 1;
 563
 564                 /* delete from next_id on up */
 565                 oa->o_valid |= OBD_MD_FLID;
 566                 oa->o_id = oscc->oscc_next_id - 1;
 567
 568                 rc = osc_real_create(exp, oa, ea, NULL);
 569
 570                 cfs_spin_lock(&oscc->oscc_lock);
 571                 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
 572                 if (rc == 0 || rc == -ENOSPC) {
 573                         struct obd_connect_data *ocd;
 574
 575                         if (rc == -ENOSPC)
 576                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 577                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 578
 579                         oscc->oscc_last_id = oa->o_id;
 580                         ocd = &imp->imp_connect_data;
 581                         if (ocd->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN) {
 582                                 CDEBUG(D_HA, "%s: Skip orphan set, reset last "
 583                                        "objid\n", oscc->oscc_obd->obd_name);
 584                                 oscc->oscc_next_id = oa->o_id + 1;
 585                         }
 586
 587                         /* sanity check for next objid. see bug 17025 */
 588                         LASSERT(oscc->oscc_next_id == oa->o_id + 1);
 589
 590                         CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
 591                                LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
 592                                oscc->oscc_last_id, rc);
 593                 } else {
 594                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 595                                oscc->oscc_obd->obd_name, rc);
 596                 }
 597
 598                 cfs_waitq_signal(&oscc->oscc_waitq);
 599                 cfs_spin_unlock(&oscc->oscc_lock);
 600
 601                 if (rc < 0)
 602                         RETURN(rc);
 603         }
 604
 605         lsm = *ea;
 606         if (lsm == NULL) {
 607                 rc = obd_alloc_memmd(exp, &lsm);
 608                 if (rc < 0)
 609                         RETURN(rc);
 610         }
 611
 612         while (1) {
 613                 if (oscc_in_sync(oscc))
 614                         CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
 615                                oscc->oscc_obd->obd_name);
 616
 617                 rc = oscc_precreate(oscc);
 618                 if (rc)
 619                         CDEBUG(D_HA,"%s: error create %d\n",
 620                                oscc->oscc_obd->obd_name, rc);
 621
 622                 cfs_spin_lock(&oscc->oscc_lock);
 623
 624                 /* wakeup but recovery did not finished */
 625                 if ((oscc->oscc_obd->u.cli.cl_import->imp_invalid) ||
 626                     (oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 627                         rc = -EIO;
 628                         cfs_spin_unlock(&oscc->oscc_lock);
 629                         break;
 630                 }
 631
 632                 if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 633                         rc = -ENOSPC;
 634                         cfs_spin_unlock(&oscc->oscc_lock);
 635                         break;
 636                 }
 637
 638                 if (oscc->oscc_flags & OSCC_FLAG_RDONLY) {
 639                         rc = -EROFS;
 640                         cfs_spin_unlock(&oscc->oscc_lock);
 641                         break;
 642                 }
 643
 644                 // Should we report -EIO error ?
 645                 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
 646                         cfs_spin_unlock(&oscc->oscc_lock);
 647                         break;
 648                 }
 649
 650                 /**
 651                  * If this is DELORPHAN process, no need create object here,
 652                  * otherwise this will create a gap of object id, and MDS
 653                  * might create some orphan log (mds_lov_update_objids), then
 654                  * remove objects wrongly on OST. Bug 21379.
 655                  */
 656                 if (oa->o_valid & OBD_MD_FLFLAGS &&
 657                         oa->o_flags == OBD_FL_DELORPHAN) {
 658                         cfs_spin_unlock(&oscc->oscc_lock);
 659                         break;
 660                 }
 661
 662                 if (oscc_has_objects_nolock(oscc, 1)) {
 663                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 664                         oa->o_id = oscc->oscc_next_id;
 665                         lsm->lsm_object_id = oscc->oscc_next_id;
 666                         *ea = lsm;
 667                         oscc->oscc_next_id++;
 668                         cfs_spin_unlock(&oscc->oscc_lock);
 669
 670                         CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
 671                                exp->exp_obd->obd_name, oscc->oscc_next_id);
 672                         break;
 673                 }
 674
 675                 cfs_spin_unlock(&oscc->oscc_lock);
 676         }
 677
 678         if (rc == 0) {
 679                 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
 680                        obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
 681         } else {
 682                 if (*ea == NULL)
 683                         obd_free_memmd(exp, &lsm);
 684                 if (del_orphan != 0 && rc != -EIO)
 685                         /* Ignore non-IO precreate error for clear orphan */
 686                         rc = 0;
 687         }
 688         RETURN(rc);
 689 }
 690
 691 void oscc_init(struct obd_device *obd)
 692 {
 693         struct osc_creator *oscc;
 694
 695         if (obd == NULL)
 696                 return;
 697
 698         oscc = &obd->u.cli.cl_oscc;
 699
 700         memset(oscc, 0, sizeof(*oscc));
 701
 702         cfs_waitq_init(&oscc->oscc_waitq);
 703         cfs_spin_lock_init(&oscc->oscc_lock);
 704         oscc->oscc_obd = obd;
 705         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 706         oscc->oscc_max_grow_count = OST_MAX_PRECREATE;
 707
 708         oscc->oscc_next_id = 2;
 709         oscc->oscc_last_id = 1;
 710         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 711
 712         CFS_INIT_LIST_HEAD(&oscc->oscc_wait_create_list);
 713
 714         /* XXX the export handle should give the oscc the last object */
 715         /* oed->oed_oscc.oscc_last_id = exph->....; */
 716 }
 717
 718 void oscc_fini(struct obd_device *obd)
 719 {
 720         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
 721         ENTRY;
 722
 723
 724         cfs_spin_lock(&oscc->oscc_lock);
 725         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 726         oscc->oscc_flags |= OSCC_FLAG_EXITING;
 727         cfs_spin_unlock(&oscc->oscc_lock);
 728 }