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  * Copyright (c) 2011, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * lustre/osc/osc_create.c
  39  * For testing and management it is treated as an obd_device,
  40  * although * it does not export a full OBD method table (the
  41  * requests are coming * in over the wire, so object target modules
  42  * do not have a full * method table.)
  43  *
  44  * Author: Peter Braam <braam@clusterfs.com>
  45  */
  46
  47 #ifndef EXPORT_SYMTAB
  48 # define EXPORT_SYMTAB
  49 #endif
  50 #define DEBUG_SUBSYSTEM S_OSC
  51
  52 #ifdef __KERNEL__
  53 # include <libcfs/libcfs.h>
  54 #else /* __KERNEL__ */
  55 # include <liblustre.h>
  56 #endif
  57
  58 #ifdef  __CYGWIN__
  59 # include <ctype.h>
  60 #endif
  61
  62 #include <lustre_dlm.h>
  63 #include <obd_class.h>
  64 #include "osc_internal.h"
  65
  66 /* XXX need AT adjust ? */
  67 #define osc_create_timeout      (obd_timeout / 2)
  68
  69 struct osc_create_async_args {
  70         struct osc_creator      *rq_oscc;
  71         struct lov_stripe_md    *rq_lsm;
  72         struct obd_info         *rq_oinfo;
  73         int                      rq_grow_count;
  74 };
  75
  76 static int oscc_internal_create(struct osc_creator *oscc);
  77 static int handle_async_create(struct ptlrpc_request *req, int rc);
  78
  79 static int osc_interpret_create(const struct lu_env *env,
  80                                 struct ptlrpc_request *req, void *data, int rc)
  81 {
  82         struct osc_create_async_args *args = ptlrpc_req_async_args(req);
  83         struct osc_creator *oscc = args->rq_oscc;
  84         struct ost_body *body = NULL;
  85         struct ptlrpc_request *fake_req, *pos;
  86         ENTRY;
  87
  88         if (req->rq_repmsg) {
  89                 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
  90                 if (body == NULL && rc == 0)
  91                         rc = -EPROTO;
  92         }
  93
  94         LASSERT(oscc && (oscc->oscc_obd != LP_POISON));
  95
  96         cfs_spin_lock(&oscc->oscc_lock);
  97         oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
  98         switch (rc) {
  99         case 0: {
 100                 if (body) {
 101                         int diff =ostid_id(&body->oa.o_oi)- oscc->oscc_last_id;
 102
 103                         /* oscc_internal_create() stores the original value of
 104                          * grow_count in osc_create_async_args::rq_grow_count.
 105                          * We can't compare against oscc_grow_count directly,
 106                          * because it may have been increased while the RPC
 107                          * is in flight, so we would always find ourselves
 108                          * having created fewer objects and decreasing the
 109                          * precreate request size.  b=18577 */
 110                         if (diff < args->rq_grow_count) {
 111                                 /* the OST has not managed to create all the
 112                                  * objects we asked for */
 113                                 oscc->oscc_grow_count = max(diff,
 114                                                             OST_MIN_PRECREATE);
 115                                 /* don't bump grow_count next time */
 116                                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 117                         } else {
 118                                 /* the OST is able to keep up with the work,
 119                                  * we could consider increasing grow_count
 120                                  * next time if needed */
 121                                 oscc->oscc_flags &= ~OSCC_FLAG_LOW;
 122                         }
 123                         oscc->oscc_last_id = ostid_id(&body->oa.o_oi);
 124                 }
 125                 cfs_spin_unlock(&oscc->oscc_lock);
 126                 break;
 127         }
 128         case -EROFS:
 129                 oscc->oscc_flags |= OSCC_FLAG_RDONLY;
 130         case -ENOSPC:
 131         case -EFBIG:
 132                 if (rc != -EROFS) {
 133                         oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 134                         if (body && rc == -ENOSPC) {
 135                                 oscc->oscc_last_id = body->oa.o_id;
 136                                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
 137
 138                                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
 139                                     (body->oa.o_flags & OBD_FL_NOSPC_BLK))
 140                                         oscc->oscc_flags |= OSCC_FLAG_NOSPC_BLK;
 141                                 else
 142                                         rc = 0;
 143                         }
 144                 }
 145                 cfs_spin_unlock(&oscc->oscc_lock);
 146                 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
 147                 break;
 148         case -EIO: {
 149                 /* filter always set body->oa.o_id as the last_id
 150                  * of filter (see filter_handle_precreate for detail)*/
 151                 if (body && body->oa.o_id > oscc->oscc_last_id)
 152                         oscc->oscc_last_id = body->oa.o_id;
 153                 cfs_spin_unlock(&oscc->oscc_lock);
 154                 break;
 155         }
 156         case -EINTR:
 157         case -EWOULDBLOCK: {
 158                 /* aka EAGAIN we should not delay create if import failed -
 159                  * this avoid client stick in create and avoid race with
 160                  * delorphan */
 161                 /* EINTR say - old create request is killed due mds<>ost
 162                  * eviction - OSCC_FLAG_RECOVERING can already set due
 163                  * IMP_DISCONN event */
 164                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 165                 /* oscc->oscc_grow_count = OST_MIN_PRECREATE; */
 166                 cfs_spin_unlock(&oscc->oscc_lock);
 167                 break;
 168         }
 169         default: {
 170                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 171                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
 172                 cfs_spin_unlock(&oscc->oscc_lock);
 173                 DEBUG_REQ(D_ERROR, req,
 174                           "Unknown rc %d from async create: failing oscc", rc);
 175                 ptlrpc_fail_import(req->rq_import,
 176                                    lustre_msg_get_conn_cnt(req->rq_reqmsg));
 177         }
 178         }
 179
 180         CDEBUG(D_HA, "preallocated through id "LPU64" (next to use "LPU64")\n",
 181                oscc->oscc_last_id, oscc->oscc_next_id);
 182
 183         cfs_spin_lock(&oscc->oscc_lock);
 184         cfs_list_for_each_entry_safe(fake_req, pos,
 185                                      &oscc->oscc_wait_create_list, rq_list) {
 186                 if (handle_async_create(fake_req, rc) == -EAGAIN) {
 187                         oscc_internal_create(oscc);
 188                         /* sending request should be never fail because
 189                          * osc use preallocated requests pool */
 190                         GOTO(exit_wakeup, rc);
 191                 }
 192         }
 193         cfs_spin_unlock(&oscc->oscc_lock);
 194
 195 exit_wakeup:
 196         cfs_waitq_signal(&oscc->oscc_waitq);
 197         RETURN(rc);
 198 }
 199
 200 static int oscc_internal_create(struct osc_creator *oscc)
 201 {
 202         struct osc_create_async_args *args;
 203         struct ptlrpc_request *request;
 204         struct ost_body *body;
 205         ENTRY;
 206
 207         LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
 208
 209         /* Do not check for a degraded OST here - bug21563/bug18539 */
 210         if (oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 211                 cfs_spin_unlock(&oscc->oscc_lock);
 212                 RETURN(0);
 213         }
 214
 215         /* we need check it before OSCC_FLAG_CREATING - because need
 216          * see lower number of precreate objects */
 217         if (oscc->oscc_grow_count < oscc->oscc_max_grow_count &&
 218             ((oscc->oscc_flags & OSCC_FLAG_LOW) == 0) &&
 219             (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
 220                    (oscc->oscc_grow_count / 4 + 1)) {
 221                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 222                 oscc->oscc_grow_count *= 2;
 223         }
 224
 225         if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
 226                 cfs_spin_unlock(&oscc->oscc_lock);
 227                 RETURN(0);
 228         }
 229
 230         if (oscc->oscc_grow_count > oscc->oscc_max_grow_count / 2)
 231                 oscc->oscc_grow_count = oscc->oscc_max_grow_count / 2;
 232
 233         oscc->oscc_flags |= OSCC_FLAG_CREATING;
 234         cfs_spin_unlock(&oscc->oscc_lock);
 235
 236         request = ptlrpc_request_alloc_pack(oscc->oscc_obd->u.cli.cl_import,
 237                                             &RQF_OST_CREATE,
 238                                             LUSTRE_OST_VERSION, OST_CREATE);
 239         if (request == NULL) {
 240                 cfs_spin_lock(&oscc->oscc_lock);
 241                 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
 242                 cfs_spin_unlock(&oscc->oscc_lock);
 243                 RETURN(-ENOMEM);
 244         }
 245
 246         request->rq_request_portal = OST_CREATE_PORTAL;
 247         ptlrpc_at_set_req_timeout(request);
 248         body = req_capsule_client_get(&request->rq_pill, &RMF_OST_BODY);
 249         args = ptlrpc_req_async_args(request);
 250         args->rq_oscc = oscc;
 251
 252         cfs_spin_lock(&oscc->oscc_lock);
 253         args->rq_grow_count = oscc->oscc_grow_count;
 254
 255         if (likely(fid_seq_is_mdt(oscc->oscc_oa.o_seq))) {
 256                 body->oa.o_oi.oi_seq = oscc->oscc_oa.o_seq;
 257                 body->oa.o_oi.oi_id  = oscc->oscc_last_id +
 258                                        oscc->oscc_grow_count;
 259         } else {
 260                 /*Just warning here currently, since not sure how fid-on-ost
 261                  *will be implemented here */
 262                 CWARN("o_seq: "LPU64" is not indicate any MDTs.\n",
 263                        oscc->oscc_oa.o_seq);
 264         }
 265         cfs_spin_unlock(&oscc->oscc_lock);
 266
 267         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 268         CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
 269                body->oa.o_id, oscc->oscc_last_id);
 270
 271         /* we should not resend create request - anyway we will have delorphan
 272          * and kill these objects */
 273         request->rq_no_delay = request->rq_no_resend = 1;
 274         ptlrpc_request_set_replen(request);
 275
 276         request->rq_interpret_reply = osc_interpret_create;
 277         ptlrpcd_add_req(request, PDL_POLICY_ROUND, -1);
 278
 279         RETURN(0);
 280 }
 281
 282 static int oscc_has_objects_nolock(struct osc_creator *oscc, int count)
 283 {
 284         return ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 285 }
 286
 287 static int oscc_has_objects(struct osc_creator *oscc, int count)
 288 {
 289         int have_objs;
 290
 291         cfs_spin_lock(&oscc->oscc_lock);
 292         have_objs = oscc_has_objects_nolock(oscc, count);
 293         cfs_spin_unlock(&oscc->oscc_lock);
 294
 295         return have_objs;
 296 }
 297
 298 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 299 {
 300         int have_objs;
 301         int ost_unusable;
 302
 303         ost_unusable = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 304
 305         cfs_spin_lock(&oscc->oscc_lock);
 306         ost_unusable |= (OSCC_FLAG_NOSPC | OSCC_FLAG_RDONLY |
 307                          OSCC_FLAG_EXITING) & oscc->oscc_flags;
 308         have_objs = oscc_has_objects_nolock(oscc, count);
 309
 310         if (!ost_unusable && !have_objs)
 311                 /* they release lock himself */
 312                 have_objs = oscc_internal_create(oscc);
 313         else
 314                 cfs_spin_unlock(&oscc->oscc_lock);
 315
 316         return have_objs || ost_unusable;
 317 }
 318
 319 static int oscc_precreate(struct osc_creator *oscc)
 320 {
 321         struct l_wait_info lwi;
 322         int rc = 0;
 323         ENTRY;
 324
 325         if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
 326                 RETURN(0);
 327
 328         /* we should be not block forever - because client's create rpc can
 329          * stick in mds for long time and forbid client reconnect */
 330         lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(osc_create_timeout)),
 331                           NULL, NULL);
 332
 333         rc = l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 334         RETURN(rc);
 335 }
 336
 337 static int oscc_in_sync(struct osc_creator *oscc)
 338 {
 339         int sync;
 340
 341         cfs_spin_lock(&oscc->oscc_lock);
 342         sync = oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS;
 343         cfs_spin_unlock(&oscc->oscc_lock);
 344
 345         return sync;
 346 }
 347
 348 /* decide if the OST has remaining object, return value :
 349         0 : the OST has remaining objects, may or may not send precreation RPC.
 350         1 : the OST has no remaining object, and the sent precreation RPC
 351             has not been completed yet.
 352         2 : the OST has no remaining object, and will not get any for
 353             a potentially very long time
 354      1000 : unusable
 355  */
 356 int osc_precreate(struct obd_export *exp)
 357 {
 358         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 359         struct obd_import *imp = exp->exp_imp_reverse;
 360         int rc;
 361         ENTRY;
 362
 363         LASSERT(oscc != NULL);
 364         if (imp != NULL && imp->imp_deactive)
 365                 GOTO(out_nolock, rc = 1000);
 366
 367         /* Handle critical states first */
 368         cfs_spin_lock(&oscc->oscc_lock);
 369         if (oscc->oscc_flags & OSCC_FLAG_NOSPC_BLK ||
 370             oscc->oscc_flags & OSCC_FLAG_RDONLY ||
 371             oscc->oscc_flags & OSCC_FLAG_EXITING)
 372                 GOTO(out, rc = 1000);
 373
 374         if ((oscc->oscc_flags & OSCC_FLAG_RECOVERING) ||
 375             (oscc->oscc_flags & OSCC_FLAG_DEGRADED))
 376                 GOTO(out, rc = 2);
 377
 378         if (oscc_has_objects_nolock(oscc, oscc->oscc_grow_count / 2))
 379                 GOTO(out, rc = 0);
 380
 381         /* Return 0, if we have at least one object - bug 22884 */
 382         rc = oscc_has_objects_nolock(oscc, 1) ? 0 : 1;
 383
 384         if (oscc->oscc_flags & OSCC_FLAG_NOSPC)
 385                 GOTO(out, (rc == 0) ? 0 : 1000);
 386
 387         /* Do not check for OSCC_FLAG_CREATING flag here, let
 388          * osc_precreate() call oscc_internal_create() and
 389          * adjust oscc_grow_count bug21563 */
 390         if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS)
 391                 GOTO(out, rc);
 392
 393         if (oscc_internal_create(oscc))
 394                 GOTO(out_nolock, rc = 1000);
 395
 396         RETURN(rc);
 397 out:
 398         cfs_spin_unlock(&oscc->oscc_lock);
 399 out_nolock:
 400         return rc;
 401 }
 402
 403 static int handle_async_create(struct ptlrpc_request *req, int rc)
 404 {
 405         struct osc_create_async_args *args = ptlrpc_req_async_args(req);
 406         struct osc_creator    *oscc = args->rq_oscc;
 407         struct lov_stripe_md  *lsm  = args->rq_lsm;
 408         struct obd_info       *oinfo = args->rq_oinfo;
 409         struct obdo           *oa = oinfo->oi_oa;
 410
 411         LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
 412
 413         if(rc)
 414                 GOTO(out_wake, rc);
 415
 416         /* Handle the critical type errors first.
 417          * Should we also test cl_import state as well ? */
 418         if (oscc->oscc_flags & OSCC_FLAG_EXITING)
 419                 GOTO(out_wake, rc = -EIO);
 420
 421         if (oscc->oscc_flags & OSCC_FLAG_NOSPC_BLK)
 422                 GOTO(out_wake, rc = -ENOSPC);
 423
 424         if (oscc->oscc_flags & OSCC_FLAG_RDONLY)
 425                 GOTO(out_wake, rc = -EROFS);
 426
 427         /* should be try wait until recovery finished */
 428         if((oscc->oscc_flags & OSCC_FLAG_RECOVERING) ||
 429            (oscc->oscc_flags & OSCC_FLAG_DEGRADED))
 430                 RETURN(-EAGAIN);
 431
 432         if (oscc_has_objects_nolock(oscc, 1)) {
 433                 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 434                 oa->o_id = oscc->oscc_next_id;
 435                 lsm->lsm_object_id = oscc->oscc_next_id;
 436                 oscc->oscc_next_id++;
 437
 438                 CDEBUG(D_RPCTRACE, " set oscc_next_id = "LPU64"\n",
 439                        oscc->oscc_next_id);
 440                 GOTO(out_wake, rc = 0);
 441         }
 442
 443         /* we don't have objects now - continue wait */
 444         RETURN(-EAGAIN);
 445
 446 out_wake:
 447
 448         rc = oinfo->oi_cb_up(oinfo, rc);
 449         ptlrpc_fakereq_finished(req);
 450
 451         RETURN(rc);
 452 }
 453
 454 static int async_create_interpret(const struct lu_env *env,
 455                                   struct ptlrpc_request *req, void *data,
 456                                   int rc)
 457 {
 458         struct osc_create_async_args *args = ptlrpc_req_async_args(req);
 459         struct osc_creator    *oscc = args->rq_oscc;
 460         int ret;
 461
 462         cfs_spin_lock(&oscc->oscc_lock);
 463         ret = handle_async_create(req, rc);
 464         cfs_spin_unlock(&oscc->oscc_lock);
 465
 466         return ret;
 467 }
 468
 469 int osc_create_async(struct obd_export *exp, struct obd_info *oinfo,
 470                      struct lov_stripe_md **ea, struct obd_trans_info *oti)
 471 {
 472         int rc;
 473         struct ptlrpc_request *fake_req;
 474         struct osc_create_async_args *args;
 475         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 476         struct obdo *oa = oinfo->oi_oa;
 477         ENTRY;
 478
 479         if ((oa->o_valid & OBD_MD_FLGROUP) && !fid_seq_is_mdt(oa->o_seq)) {
 480                 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
 481                 rc = oinfo->oi_cb_up(oinfo, rc);
 482                 RETURN(rc);
 483         }
 484
 485         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 486             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 487                 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
 488                 rc = oinfo->oi_cb_up(oinfo, rc);
 489                 RETURN(rc);
 490         }
 491
 492         LASSERT((*ea) != NULL);
 493
 494         fake_req = ptlrpc_prep_fakereq(oscc->oscc_obd->u.cli.cl_import,
 495                                        osc_create_timeout,
 496                                        async_create_interpret);
 497         if (fake_req == NULL) {
 498                 rc = oinfo->oi_cb_up(oinfo, -ENOMEM);
 499                 RETURN(-ENOMEM);
 500         }
 501
 502         args = ptlrpc_req_async_args(fake_req);
 503         CLASSERT(sizeof(*args) <= sizeof(fake_req->rq_async_args));
 504
 505         args->rq_oscc  = oscc;
 506         args->rq_lsm   = *ea;
 507         args->rq_oinfo = oinfo;
 508
 509         cfs_spin_lock(&oscc->oscc_lock);
 510         /* try fast path */
 511         rc = handle_async_create(fake_req, 0);
 512         if (rc == -EAGAIN) {
 513                 /* We don't have any objects, wait until we get a reply. */
 514                 ptlrpcd_add_req(fake_req, PDL_POLICY_ROUND, -1);
 515                 cfs_list_add(&fake_req->rq_list,
 516                              &oscc->oscc_wait_create_list);
 517                 cfs_spin_unlock(&oscc->oscc_lock);
 518                 /* EAGAIN mean - request is delayed */
 519                 rc = 0;
 520         } else {
 521                 cfs_spin_unlock(&oscc->oscc_lock);
 522                 /* need free request if was error hit or
 523                  * objects already allocated */
 524                 ptlrpc_req_finished(fake_req);
 525         }
 526
 527         RETURN(rc);
 528 }
 529
 530 int osc_create(struct obd_export *exp, struct obdo *oa,
 531                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 532 {
 533         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 534         struct obd_import  *imp  = exp->exp_obd->u.cli.cl_import;
 535         struct lov_stripe_md *lsm;
 536         int del_orphan = 0, rc = 0;
 537         ENTRY;
 538
 539         LASSERT(oa);
 540         LASSERT(ea);
 541         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 542
 543         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 544             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 545                 RETURN(osc_real_create(exp, oa, ea, oti));
 546         }
 547
 548         if (!fid_seq_is_mdt(oa->o_seq))
 549                 RETURN(osc_real_create(exp, oa, ea, oti));
 550
 551         /* this is the special case where create removes orphans */
 552         if (oa->o_valid & OBD_MD_FLFLAGS &&
 553             oa->o_flags == OBD_FL_DELORPHAN) {
 554                 cfs_spin_lock(&oscc->oscc_lock);
 555                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 556                         cfs_spin_unlock(&oscc->oscc_lock);
 557                         RETURN(-EBUSY);
 558                 }
 559                 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 560                         cfs_spin_unlock(&oscc->oscc_lock);
 561                         RETURN(0);
 562                 }
 563
 564                 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
 565                 /* seting flag LOW we prevent extra grow precreate size
 566                  * and enforce use last assigned size */
 567                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 568                 cfs_spin_unlock(&oscc->oscc_lock);
 569                 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
 570                        oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
 571
 572                 del_orphan = 1;
 573
 574                 /* delete from next_id on up */
 575                 oa->o_valid |= OBD_MD_FLID;
 576                 oa->o_id = oscc->oscc_next_id - 1;
 577
 578                 rc = osc_real_create(exp, oa, ea, NULL);
 579
 580                 cfs_spin_lock(&oscc->oscc_lock);
 581                 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
 582                 if (rc == 0 || rc == -ENOSPC) {
 583                         struct obd_connect_data *ocd;
 584
 585                         if (rc == -ENOSPC) {
 586                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 587                                 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 588                                     (oa->o_flags & OBD_FL_NOSPC_BLK))
 589                                         oscc->oscc_flags |= OSCC_FLAG_NOSPC_BLK;
 590                         }
 591                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 592
 593                         oscc->oscc_last_id = oa->o_id;
 594                         ocd = &imp->imp_connect_data;
 595                         if (ocd->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN) {
 596                                 /*
 597                                  * The OST reports back in oa->o_id from where
 598                                  * we should restart in order to skip orphan
 599                                  * objects
 600                                  */
 601                                 CDEBUG(D_HA, "%s: Skip orphan set, reset last "
 602                                        "objid\n", oscc->oscc_obd->obd_name);
 603                                 oscc->oscc_next_id = oa->o_id + 1;
 604                         }
 605
 606                         /* sanity check for next objid. see bug 17025 */
 607                         LASSERT(oscc->oscc_next_id == oa->o_id + 1);
 608
 609                         CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
 610                                LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
 611                                oscc->oscc_last_id, rc);
 612                 } else {
 613                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 614                                oscc->oscc_obd->obd_name, rc);
 615                 }
 616
 617                 cfs_waitq_signal(&oscc->oscc_waitq);
 618                 cfs_spin_unlock(&oscc->oscc_lock);
 619
 620                 if (rc < 0)
 621                         RETURN(rc);
 622         }
 623
 624         lsm = *ea;
 625         if (lsm == NULL) {
 626                 rc = obd_alloc_memmd(exp, &lsm);
 627                 if (rc < 0)
 628                         RETURN(rc);
 629         }
 630
 631         while (1) {
 632                 if (oscc_in_sync(oscc))
 633                         CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
 634                                oscc->oscc_obd->obd_name);
 635
 636                 rc = oscc_precreate(oscc);
 637                 if (rc)
 638                         CDEBUG(D_HA,"%s: error create %d\n",
 639                                oscc->oscc_obd->obd_name, rc);
 640
 641                 cfs_spin_lock(&oscc->oscc_lock);
 642
 643                 /* wakeup but recovery did not finished */
 644                 if ((oscc->oscc_obd->u.cli.cl_import->imp_invalid) ||
 645                     (oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 646                         rc = -EIO;
 647                         cfs_spin_unlock(&oscc->oscc_lock);
 648                         break;
 649                 }
 650
 651                 if (oscc->oscc_flags & OSCC_FLAG_NOSPC_BLK) {
 652                         rc = -ENOSPC;
 653                         cfs_spin_unlock(&oscc->oscc_lock);
 654                         break;
 655                 }
 656
 657                 if (oscc->oscc_flags & OSCC_FLAG_RDONLY) {
 658                         rc = -EROFS;
 659                         cfs_spin_unlock(&oscc->oscc_lock);
 660                         break;
 661                 }
 662
 663                 // Should we report -EIO error ?
 664                 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
 665                         cfs_spin_unlock(&oscc->oscc_lock);
 666                         break;
 667                 }
 668
 669                 /**
 670                  * If this is DELORPHAN process, no need create object here,
 671                  * otherwise this will create a gap of object id, and MDS
 672                  * might create some orphan log (mds_lov_update_objids), then
 673                  * remove objects wrongly on OST. Bug 21379.
 674                  */
 675                 if (oa->o_valid & OBD_MD_FLFLAGS &&
 676                         oa->o_flags == OBD_FL_DELORPHAN) {
 677                         cfs_spin_unlock(&oscc->oscc_lock);
 678                         break;
 679                 }
 680
 681                 if (oscc_has_objects_nolock(oscc, 1)) {
 682                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 683                         oa->o_id = oscc->oscc_next_id;
 684                         lsm->lsm_object_id = oscc->oscc_next_id;
 685                         *ea = lsm;
 686                         oscc->oscc_next_id++;
 687                         cfs_spin_unlock(&oscc->oscc_lock);
 688
 689                         CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
 690                                exp->exp_obd->obd_name, oscc->oscc_next_id);
 691                         break;
 692                 }
 693
 694                 if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 695                         rc = -ENOSPC;
 696                         cfs_spin_unlock(&oscc->oscc_lock);
 697                         break;
 698                 }
 699
 700                 cfs_spin_unlock(&oscc->oscc_lock);
 701         }
 702
 703         if (rc == 0) {
 704                 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
 705                        obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
 706         } else {
 707                 if (*ea == NULL)
 708                         obd_free_memmd(exp, &lsm);
 709                 if (del_orphan != 0 && rc != -EIO)
 710                         /* Ignore non-IO precreate error for clear orphan */
 711                         rc = 0;
 712         }
 713         RETURN(rc);
 714 }
 715
 716 void oscc_init(struct obd_device *obd)
 717 {
 718         struct osc_creator *oscc;
 719
 720         if (obd == NULL)
 721                 return;
 722
 723         oscc = &obd->u.cli.cl_oscc;
 724
 725         memset(oscc, 0, sizeof(*oscc));
 726
 727         cfs_waitq_init(&oscc->oscc_waitq);
 728         cfs_spin_lock_init(&oscc->oscc_lock);
 729         oscc->oscc_obd = obd;
 730         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 731         oscc->oscc_max_grow_count = OST_MAX_PRECREATE;
 732
 733         oscc->oscc_next_id = 2;
 734         oscc->oscc_last_id = 1;
 735         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 736
 737         CFS_INIT_LIST_HEAD(&oscc->oscc_wait_create_list);
 738
 739         /* XXX the export handle should give the oscc the last object */
 740         /* oed->oed_oscc.oscc_last_id = exph->....; */
 741 }
 742
 743 void oscc_fini(struct obd_device *obd)
 744 {
 745         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
 746         ENTRY;
 747
 748
 749         cfs_spin_lock(&oscc->oscc_lock);
 750         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 751         oscc->oscc_flags |= OSCC_FLAG_EXITING;
 752         cfs_spin_unlock(&oscc->oscc_lock);
 753 }