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