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         int                      rq_grow_count;
  72 };
  73
  74 static int oscc_internal_create(struct osc_creator *oscc);
  75 static int handle_async_create(struct ptlrpc_request *req, int rc);
  76
  77 static int osc_interpret_create(const struct lu_env *env,
  78                                 struct ptlrpc_request *req, void *data, int rc)
  79 {
  80         struct osc_create_async_args *args = ptlrpc_req_async_args(req);
  81         struct osc_creator *oscc = args->rq_oscc;
  82         struct ost_body *body = NULL;
  83         struct ptlrpc_request *fake_req, *pos;
  84         ENTRY;
  85
  86         if (req->rq_repmsg) {
  87                 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
  88                 if (body == NULL && rc == 0)
  89                         rc = -EPROTO;
  90         }
  91
  92         LASSERT(oscc && (oscc->oscc_obd != LP_POISON));
  93
  94         cfs_spin_lock(&oscc->oscc_lock);
  95         oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
  96         switch (rc) {
  97         case 0: {
  98                 if (body) {
  99                         int diff =ostid_id(&body->oa.o_oi)- oscc->oscc_last_id;
 100
 101                         /* oscc_internal_create() stores the original value of
 102                          * grow_count in osc_create_async_args::rq_grow_count.
 103                          * We can't compare against oscc_grow_count directly,
 104                          * because it may have been increased while the RPC
 105                          * is in flight, so we would always find ourselves
 106                          * having created fewer objects and decreasing the
 107                          * precreate request size.  b=18577 */
 108                         if (diff < args->rq_grow_count) {
 109                                 /* the OST has not managed to create all the
 110                                  * objects we asked for */
 111                                 oscc->oscc_grow_count = max(diff,
 112                                                             OST_MIN_PRECREATE);
 113                                 /* don't bump grow_count next time */
 114                                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 115                         } else {
 116                                 /* the OST is able to keep up with the work,
 117                                  * we could consider increasing grow_count
 118                                  * next time if needed */
 119                                 oscc->oscc_flags &= ~OSCC_FLAG_LOW;
 120                         }
 121                         oscc->oscc_last_id = ostid_id(&body->oa.o_oi);
 122                 }
 123                 cfs_spin_unlock(&oscc->oscc_lock);
 124                 break;
 125         }
 126         case -EROFS:
 127                 oscc->oscc_flags |= OSCC_FLAG_RDONLY;
 128         case -ENOSPC:
 129         case -EFBIG:
 130                 if (rc != -EROFS) {
 131                         oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 132                         if (body && rc == -ENOSPC) {
 133                                 oscc->oscc_last_id = body->oa.o_id;
 134                                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
 135
 136                                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
 137                                     (body->oa.o_flags & OBD_FL_NOSPC_BLK))
 138                                         oscc->oscc_flags |= OSCC_FLAG_NOSPC_BLK;
 139                                 else
 140                                         rc = 0;
 141                         }
 142                 }
 143                 cfs_spin_unlock(&oscc->oscc_lock);
 144                 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
 145                 break;
 146         case -EIO: {
 147                 /* filter always set body->oa.o_id as the last_id
 148                  * of filter (see filter_handle_precreate for detail)*/
 149                 if (body && body->oa.o_id > oscc->oscc_last_id)
 150                         oscc->oscc_last_id = body->oa.o_id;
 151                 cfs_spin_unlock(&oscc->oscc_lock);
 152                 break;
 153         }
 154         case -EINTR:
 155         case -EWOULDBLOCK: {
 156                 /* aka EAGAIN we should not delay create if import failed -
 157                  * this avoid client stick in create and avoid race with
 158                  * delorphan */
 159                 /* EINTR say - old create request is killed due mds<>ost
 160                  * eviction - OSCC_FLAG_RECOVERING can already set due
 161                  * IMP_DISCONN event */
 162                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 163                 /* oscc->oscc_grow_count = OST_MIN_PRECREATE; */
 164                 cfs_spin_unlock(&oscc->oscc_lock);
 165                 break;
 166         }
 167         default: {
 168                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 169                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
 170                 cfs_spin_unlock(&oscc->oscc_lock);
 171                 DEBUG_REQ(D_ERROR, req,
 172                           "Unknown rc %d from async create: failing oscc", rc);
 173                 ptlrpc_fail_import(req->rq_import,
 174                                    lustre_msg_get_conn_cnt(req->rq_reqmsg));
 175         }
 176         }
 177
 178         CDEBUG(D_HA, "preallocated through id "LPU64" (next to use "LPU64")\n",
 179                oscc->oscc_last_id, oscc->oscc_next_id);
 180
 181         cfs_spin_lock(&oscc->oscc_lock);
 182         cfs_list_for_each_entry_safe(fake_req, pos,
 183                                      &oscc->oscc_wait_create_list, rq_list) {
 184                 if (handle_async_create(fake_req, rc)  == -EAGAIN) {
 185                         oscc_internal_create(oscc);
 186                         /* sending request should be never fail because
 187                          * osc use preallocated requests pool */
 188                         GOTO(exit_wakeup, rc);
 189                 }
 190         }
 191         cfs_spin_unlock(&oscc->oscc_lock);
 192
 193 exit_wakeup:
 194         cfs_waitq_signal(&oscc->oscc_waitq);
 195         RETURN(rc);
 196 }
 197
 198 static int oscc_internal_create(struct osc_creator *oscc)
 199 {
 200         struct osc_create_async_args *args;
 201         struct ptlrpc_request *request;
 202         struct ost_body *body;
 203         ENTRY;
 204
 205         LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
 206
 207         /* Do not check for a degraded OST here - bug21563/bug18539 */
 208         if (oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 209                 cfs_spin_unlock(&oscc->oscc_lock);
 210                 RETURN(0);
 211         }
 212
 213         /* we need check it before OSCC_FLAG_CREATING - because need
 214          * see lower number of precreate objects */
 215         if (oscc->oscc_grow_count < oscc->oscc_max_grow_count &&
 216             ((oscc->oscc_flags & OSCC_FLAG_LOW) == 0) &&
 217             (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
 218                    (oscc->oscc_grow_count / 4 + 1)) {
 219                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 220                 oscc->oscc_grow_count *= 2;
 221         }
 222
 223         if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
 224                 cfs_spin_unlock(&oscc->oscc_lock);
 225                 RETURN(0);
 226         }
 227
 228         if (oscc->oscc_grow_count > oscc->oscc_max_grow_count / 2)
 229                 oscc->oscc_grow_count = oscc->oscc_max_grow_count / 2;
 230
 231         oscc->oscc_flags |= OSCC_FLAG_CREATING;
 232         cfs_spin_unlock(&oscc->oscc_lock);
 233
 234         request = ptlrpc_request_alloc_pack(oscc->oscc_obd->u.cli.cl_import,
 235                                             &RQF_OST_CREATE,
 236                                             LUSTRE_OST_VERSION, OST_CREATE);
 237         if (request == NULL) {
 238                 cfs_spin_lock(&oscc->oscc_lock);
 239                 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
 240                 cfs_spin_unlock(&oscc->oscc_lock);
 241                 RETURN(-ENOMEM);
 242         }
 243
 244         request->rq_request_portal = OST_CREATE_PORTAL;
 245         ptlrpc_at_set_req_timeout(request);
 246         body = req_capsule_client_get(&request->rq_pill, &RMF_OST_BODY);
 247         args = ptlrpc_req_async_args(request);
 248         args->rq_oscc = oscc;
 249
 250         cfs_spin_lock(&oscc->oscc_lock);
 251         args->rq_grow_count = oscc->oscc_grow_count;
 252
 253         if (likely(fid_seq_is_mdt(oscc->oscc_oa.o_seq))) {
 254                 body->oa.o_oi.oi_seq = oscc->oscc_oa.o_seq;
 255                 body->oa.o_oi.oi_id  = oscc->oscc_last_id +
 256                                        oscc->oscc_grow_count;
 257         } else {
 258                 /*Just warning here currently, since not sure how fid-on-ost
 259                  *will be implemented here */
 260                 CWARN("o_seq: "LPU64" is not indicate any MDTs.\n",
 261                        oscc->oscc_oa.o_seq);
 262         }
 263         cfs_spin_unlock(&oscc->oscc_lock);
 264
 265         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 266         CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
 267                body->oa.o_id, oscc->oscc_last_id);
 268
 269         /* we should not resend create request - anyway we will have delorphan
 270          * and kill these objects */
 271         request->rq_no_delay = request->rq_no_resend = 1;
 272         ptlrpc_request_set_replen(request);
 273
 274         request->rq_interpret_reply = osc_interpret_create;
 275         ptlrpcd_add_req(request, PSCOPE_OTHER);
 276
 277         RETURN(0);
 278 }
 279
 280 static int oscc_has_objects_nolock(struct osc_creator *oscc, int count)
 281 {
 282         return ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 283 }
 284
 285
 286 static int oscc_has_objects(struct osc_creator *oscc, int count)
 287 {
 288         int have_objs;
 289
 290         cfs_spin_lock(&oscc->oscc_lock);
 291         have_objs = oscc_has_objects_nolock(oscc, count);
 292         cfs_spin_unlock(&oscc->oscc_lock);
 293
 294         return have_objs;
 295 }
 296
 297 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 298 {
 299         int have_objs;
 300         int ost_unusable;
 301
 302         ost_unusable = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 303
 304         cfs_spin_lock(&oscc->oscc_lock);
 305         ost_unusable |= (OSCC_FLAG_NOSPC | OSCC_FLAG_RDONLY |
 306                          OSCC_FLAG_EXITING) & oscc->oscc_flags;
 307         have_objs = oscc_has_objects_nolock(oscc, count);
 308
 309         if (!ost_unusable && !have_objs)
 310                 /* they release lock himself */
 311                 have_objs = oscc_internal_create(oscc);
 312         else
 313                 cfs_spin_unlock(&oscc->oscc_lock);
 314
 315         return have_objs || ost_unusable;
 316 }
 317
 318 static int oscc_precreate(struct osc_creator *oscc)
 319 {
 320         struct l_wait_info lwi;
 321         int rc = 0;
 322         ENTRY;
 323
 324         if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
 325                 RETURN(0);
 326
 327         /* we should be not block forever - because client's create rpc can
 328          * stick in mds for long time and forbid client reconnect */
 329         lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(osc_create_timeout)),
 330                           NULL, NULL);
 331
 332         rc = l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 333         RETURN(rc);
 334 }
 335
 336 static int oscc_in_sync(struct osc_creator *oscc)
 337 {
 338         int sync;
 339
 340         cfs_spin_lock(&oscc->oscc_lock);
 341         sync = oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS;
 342         cfs_spin_unlock(&oscc->oscc_lock);
 343
 344         return sync;
 345 }
 346
 347 /* decide if the OST has remaining object, return value :
 348         0 : the OST has remaining objects, may or may not send precreation RPC.
 349         1 : the OST has no remaining object, and the sent precreation RPC
 350             has not been completed yet.
 351         2 : the OST has no remaining object, and will not get any for
 352             a potentially very long time
 353      1000 : unusable
 354  */
 355 int osc_precreate(struct obd_export *exp)
 356 {
 357         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 358         struct obd_import *imp = exp->exp_imp_reverse;
 359         int rc;
 360         ENTRY;
 361
 362         LASSERT(oscc != NULL);
 363         if (imp != NULL && imp->imp_deactive)
 364                 GOTO(out_nolock, rc = 1000);
 365
 366         /* Handle critical states first */
 367         cfs_spin_lock(&oscc->oscc_lock);
 368         if (oscc->oscc_flags & OSCC_FLAG_NOSPC_BLK ||
 369             oscc->oscc_flags & OSCC_FLAG_RDONLY ||
 370             oscc->oscc_flags & OSCC_FLAG_EXITING)
 371                 GOTO(out, rc = 1000);
 372
 373         if ((oscc->oscc_flags & OSCC_FLAG_RECOVERING) ||
 374             (oscc->oscc_flags & OSCC_FLAG_DEGRADED))
 375                 GOTO(out, rc = 2);
 376
 377         if (oscc_has_objects_nolock(oscc, oscc->oscc_grow_count / 2))
 378                 GOTO(out, rc = 0);
 379
 380         /* Return 0, if we have at least one object - bug 22884 */
 381         rc = oscc_has_objects_nolock(oscc, 1) ? 0 : 1;
 382
 383         if (oscc->oscc_flags & OSCC_FLAG_NOSPC)
 384                 GOTO(out, (rc == 0) ? 0 : 1000);
 385
 386         /* Do not check for OSCC_FLAG_CREATING flag here, let
 387          * osc_precreate() call oscc_internal_create() and
 388          * adjust oscc_grow_count bug21563 */
 389         if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS)
 390                 GOTO(out, rc);
 391
 392         if (oscc_internal_create(oscc))
 393                 GOTO(out_nolock, rc = 1000);
 394
 395         RETURN(rc);
 396 out:
 397         cfs_spin_unlock(&oscc->oscc_lock);
 398 out_nolock:
 399         return rc;
 400 }
 401
 402 static int handle_async_create(struct ptlrpc_request *req, int rc)
 403 {
 404         struct osc_create_async_args *args = ptlrpc_req_async_args(req);
 405         struct osc_creator    *oscc = args->rq_oscc;
 406         struct lov_stripe_md  *lsm  = args->rq_lsm;
 407         struct obd_info       *oinfo = args->rq_oinfo;
 408         struct obdo           *oa = oinfo->oi_oa;
 409
 410         LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
 411
 412         if(rc)
 413                 GOTO(out_wake, rc);
 414
 415         /* Handle the critical type errors first.
 416          * Should we also test cl_import state as well ? */
 417         if (oscc->oscc_flags & OSCC_FLAG_EXITING)
 418                 GOTO(out_wake, rc = -EIO);
 419
 420         if (oscc->oscc_flags & OSCC_FLAG_NOSPC_BLK)
 421                 GOTO(out_wake, rc = -ENOSPC);
 422
 423         if (oscc->oscc_flags & OSCC_FLAG_RDONLY)
 424                 GOTO(out_wake, rc = -EROFS);
 425
 426         /* should be try wait until recovery finished */
 427         if((oscc->oscc_flags & OSCC_FLAG_RECOVERING) ||
 428            (oscc->oscc_flags & OSCC_FLAG_DEGRADED))
 429                 RETURN(-EAGAIN);
 430
 431         if (oscc_has_objects_nolock(oscc, 1)) {
 432                 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 433                 oa->o_id = oscc->oscc_next_id;
 434                 lsm->lsm_object_id = oscc->oscc_next_id;
 435                 oscc->oscc_next_id++;
 436
 437                 CDEBUG(D_RPCTRACE, " set oscc_next_id = "LPU64"\n",
 438                        oscc->oscc_next_id);
 439                 GOTO(out_wake, rc = 0);
 440         }
 441
 442         /* we don't have objects now - continue wait */
 443         RETURN(-EAGAIN);
 444
 445 out_wake:
 446
 447         rc = oinfo->oi_cb_up(oinfo, rc);
 448         ptlrpc_fakereq_finished(req);
 449
 450         RETURN(rc);
 451 }
 452
 453 static int async_create_interpret(const struct lu_env *env,
 454                                   struct ptlrpc_request *req, void *data,
 455                                   int rc)
 456 {
 457         struct osc_create_async_args *args = ptlrpc_req_async_args(req);
 458         struct osc_creator    *oscc = args->rq_oscc;
 459         int ret;
 460
 461         cfs_spin_lock(&oscc->oscc_lock);
 462         ret = handle_async_create(req, rc);
 463         cfs_spin_unlock(&oscc->oscc_lock);
 464
 465         return ret;
 466 }
 467
 468 int osc_create_async(struct obd_export *exp, struct obd_info *oinfo,
 469                      struct lov_stripe_md **ea, struct obd_trans_info *oti)
 470 {
 471         int rc;
 472         struct ptlrpc_request *fake_req;
 473         struct osc_create_async_args *args;
 474         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 475         struct obdo *oa = oinfo->oi_oa;
 476         ENTRY;
 477
 478         if ((oa->o_valid & OBD_MD_FLGROUP) && !fid_seq_is_mdt(oa->o_seq)) {
 479                 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
 480                 rc = oinfo->oi_cb_up(oinfo, rc);
 481                 RETURN(rc);
 482         }
 483
 484         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 485             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 486                 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
 487                 rc = oinfo->oi_cb_up(oinfo, rc);
 488                 RETURN(rc);
 489         }
 490
 491         LASSERT((*ea) != NULL);
 492
 493         fake_req = ptlrpc_prep_fakereq(oscc->oscc_obd->u.cli.cl_import,
 494                                        osc_create_timeout,
 495                                        async_create_interpret);
 496         if (fake_req == NULL) {
 497                 rc = oinfo->oi_cb_up(oinfo, -ENOMEM);
 498                 RETURN(-ENOMEM);
 499         }
 500
 501         args = ptlrpc_req_async_args(fake_req);
 502         CLASSERT(sizeof(*args) <= sizeof(fake_req->rq_async_args));
 503
 504         args->rq_oscc  = oscc;
 505         args->rq_lsm   = *ea;
 506         args->rq_oinfo = oinfo;
 507
 508         cfs_spin_lock(&oscc->oscc_lock);
 509         /* try fast path */
 510         rc = handle_async_create(fake_req, 0);
 511         if (rc == -EAGAIN) {
 512                 int is_add;
 513                 /* we not have objects - try wait */
 514                 is_add = ptlrpcd_add_req(fake_req, PSCOPE_OTHER);
 515                 if (!is_add)
 516                         cfs_list_add(&fake_req->rq_list,
 517                                      &oscc->oscc_wait_create_list);
 518                 else
 519                         rc = is_add;
 520         }
 521         cfs_spin_unlock(&oscc->oscc_lock);
 522
 523         if (rc != -EAGAIN)
 524                 /* need free request if was error hit or
 525                  * objects already allocated */
 526                 ptlrpc_req_finished(fake_req);
 527         else
 528                 /* EAGAIN mean - request is delayed */
 529                 rc = 0;
 530
 531         RETURN(rc);
 532 }
 533
 534 int osc_create(struct obd_export *exp, struct obdo *oa,
 535                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 536 {
 537         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 538         struct obd_import  *imp  = exp->exp_obd->u.cli.cl_import;
 539         struct lov_stripe_md *lsm;
 540         int del_orphan = 0, rc = 0;
 541         ENTRY;
 542
 543         LASSERT(oa);
 544         LASSERT(ea);
 545         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 546
 547         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 548             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 549                 RETURN(osc_real_create(exp, oa, ea, oti));
 550         }
 551
 552         if (!fid_seq_is_mdt(oa->o_seq))
 553                 RETURN(osc_real_create(exp, oa, ea, oti));
 554
 555         /* this is the special case where create removes orphans */
 556         if (oa->o_valid & OBD_MD_FLFLAGS &&
 557             oa->o_flags == OBD_FL_DELORPHAN) {
 558                 cfs_spin_lock(&oscc->oscc_lock);
 559                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 560                         cfs_spin_unlock(&oscc->oscc_lock);
 561                         RETURN(-EBUSY);
 562                 }
 563                 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 564                         cfs_spin_unlock(&oscc->oscc_lock);
 565                         RETURN(0);
 566                 }
 567
 568                 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
 569                 /* seting flag LOW we prevent extra grow precreate size
 570                  * and enforce use last assigned size */
 571                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 572                 cfs_spin_unlock(&oscc->oscc_lock);
 573                 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
 574                        oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
 575
 576                 del_orphan = 1;
 577
 578                 /* delete from next_id on up */
 579                 oa->o_valid |= OBD_MD_FLID;
 580                 oa->o_id = oscc->oscc_next_id - 1;
 581
 582                 rc = osc_real_create(exp, oa, ea, NULL);
 583
 584                 cfs_spin_lock(&oscc->oscc_lock);
 585                 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
 586                 if (rc == 0 || rc == -ENOSPC) {
 587                         struct obd_connect_data *ocd;
 588
 589                         if (rc == -ENOSPC) {
 590                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 591                                 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 592                                     (oa->o_flags & OBD_FL_NOSPC_BLK))
 593                                         oscc->oscc_flags |= OSCC_FLAG_NOSPC_BLK;
 594                         }
 595                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 596
 597                         oscc->oscc_last_id = oa->o_id;
 598                         ocd = &imp->imp_connect_data;
 599                         if (ocd->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN) {
 600                                 /*
 601                                  * The OST reports back in oa->o_id from where
 602                                  * we should restart in order to skip orphan
 603                                  * objects
 604                                  */
 605                                 CDEBUG(D_HA, "%s: Skip orphan set, reset last "
 606                                        "objid\n", oscc->oscc_obd->obd_name);
 607                                 oscc->oscc_next_id = oa->o_id + 1;
 608                         }
 609
 610                         /* sanity check for next objid. see bug 17025 */
 611                         LASSERT(oscc->oscc_next_id == oa->o_id + 1);
 612
 613                         CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
 614                                LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
 615                                oscc->oscc_last_id, rc);
 616                 } else {
 617                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 618                                oscc->oscc_obd->obd_name, rc);
 619                 }
 620
 621                 cfs_waitq_signal(&oscc->oscc_waitq);
 622                 cfs_spin_unlock(&oscc->oscc_lock);
 623
 624                 if (rc < 0)
 625                         RETURN(rc);
 626         }
 627
 628         lsm = *ea;
 629         if (lsm == NULL) {
 630                 rc = obd_alloc_memmd(exp, &lsm);
 631                 if (rc < 0)
 632                         RETURN(rc);
 633         }
 634
 635         while (1) {
 636                 if (oscc_in_sync(oscc))
 637                         CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
 638                                oscc->oscc_obd->obd_name);
 639
 640                 rc = oscc_precreate(oscc);
 641                 if (rc)
 642                         CDEBUG(D_HA,"%s: error create %d\n",
 643                                oscc->oscc_obd->obd_name, rc);
 644
 645                 cfs_spin_lock(&oscc->oscc_lock);
 646
 647                 /* wakeup but recovery did not finished */
 648                 if ((oscc->oscc_obd->u.cli.cl_import->imp_invalid) ||
 649                     (oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 650                         rc = -EIO;
 651                         cfs_spin_unlock(&oscc->oscc_lock);
 652                         break;
 653                 }
 654
 655                 if (oscc->oscc_flags & OSCC_FLAG_NOSPC_BLK) {
 656                         rc = -ENOSPC;
 657                         cfs_spin_unlock(&oscc->oscc_lock);
 658                         break;
 659                 }
 660
 661                 if (oscc->oscc_flags & OSCC_FLAG_RDONLY) {
 662                         rc = -EROFS;
 663                         cfs_spin_unlock(&oscc->oscc_lock);
 664                         break;
 665                 }
 666
 667                 // Should we report -EIO error ?
 668                 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
 669                         cfs_spin_unlock(&oscc->oscc_lock);
 670                         break;
 671                 }
 672
 673                 /**
 674                  * If this is DELORPHAN process, no need create object here,
 675                  * otherwise this will create a gap of object id, and MDS
 676                  * might create some orphan log (mds_lov_update_objids), then
 677                  * remove objects wrongly on OST. Bug 21379.
 678                  */
 679                 if (oa->o_valid & OBD_MD_FLFLAGS &&
 680                         oa->o_flags == OBD_FL_DELORPHAN) {
 681                         cfs_spin_unlock(&oscc->oscc_lock);
 682                         break;
 683                 }
 684
 685                 if (oscc_has_objects_nolock(oscc, 1)) {
 686                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 687                         oa->o_id = oscc->oscc_next_id;
 688                         lsm->lsm_object_id = oscc->oscc_next_id;
 689                         *ea = lsm;
 690                         oscc->oscc_next_id++;
 691                         cfs_spin_unlock(&oscc->oscc_lock);
 692
 693                         CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
 694                                exp->exp_obd->obd_name, oscc->oscc_next_id);
 695                         break;
 696                 }
 697
 698                 if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 699                         rc = -ENOSPC;
 700                         cfs_spin_unlock(&oscc->oscc_lock);
 701                         break;
 702                 }
 703
 704                 cfs_spin_unlock(&oscc->oscc_lock);
 705         }
 706
 707         if (rc == 0) {
 708                 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
 709                        obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
 710         } else {
 711                 if (*ea == NULL)
 712                         obd_free_memmd(exp, &lsm);
 713                 if (del_orphan != 0 && rc != -EIO)
 714                         /* Ignore non-IO precreate error for clear orphan */
 715                         rc = 0;
 716         }
 717         RETURN(rc);
 718 }
 719
 720 void oscc_init(struct obd_device *obd)
 721 {
 722         struct osc_creator *oscc;
 723
 724         if (obd == NULL)
 725                 return;
 726
 727         oscc = &obd->u.cli.cl_oscc;
 728
 729         memset(oscc, 0, sizeof(*oscc));
 730
 731         cfs_waitq_init(&oscc->oscc_waitq);
 732         cfs_spin_lock_init(&oscc->oscc_lock);
 733         oscc->oscc_obd = obd;
 734         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 735         oscc->oscc_max_grow_count = OST_MAX_PRECREATE;
 736
 737         oscc->oscc_next_id = 2;
 738         oscc->oscc_last_id = 1;
 739         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 740
 741         CFS_INIT_LIST_HEAD(&oscc->oscc_wait_create_list);
 742
 743         /* XXX the export handle should give the oscc the last object */
 744         /* oed->oed_oscc.oscc_last_id = exph->....; */
 745 }
 746
 747 void oscc_fini(struct obd_device *obd)
 748 {
 749         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
 750         ENTRY;
 751
 752
 753         cfs_spin_lock(&oscc->oscc_lock);
 754         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 755         oscc->oscc_flags |= OSCC_FLAG_EXITING;
 756         cfs_spin_unlock(&oscc->oscc_lock);
 757 }