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 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * lustre/osc/osc_create.c
  40  * For testing and management it is treated as an obd_device,
  41  * although * it does not export a full OBD method table (the
  42  * requests are coming * in over the wire, so object target modules
  43  * do not have a full * method table.)
  44  *
  45  * Author: Peter Braam <braam@clusterfs.com>
  46  */
  47
  48 #ifndef EXPORT_SYMTAB
  49 # define EXPORT_SYMTAB
  50 #endif
  51 #define DEBUG_SUBSYSTEM S_OSC
  52
  53 #ifdef __KERNEL__
  54 # include <libcfs/libcfs.h>
  55 #else /* __KERNEL__ */
  56 # include <liblustre.h>
  57 #endif
  58
  59 #ifdef  __CYGWIN__
  60 # include <ctype.h>
  61 #endif
  62
  63 #include <lustre_dlm.h>
  64 #include <obd_class.h>
  65 #include "osc_internal.h"
  66
  67 /* XXX need AT adjust ? */
  68 #define osc_create_timeout      (obd_timeout / 2)
  69
  70 struct osc_create_async_args {
  71         struct osc_creator      *rq_oscc;
  72         struct lov_stripe_md    *rq_lsm;
  73         struct obd_info         *rq_oinfo;
  74         int                      rq_grow_count;
  75 };
  76
  77 static int oscc_internal_create(struct osc_creator *oscc);
  78 static int handle_async_create(struct ptlrpc_request *req, int rc);
  79
  80 static int osc_interpret_create(const struct lu_env *env,
  81                                 struct ptlrpc_request *req, void *data, int rc)
  82 {
  83         struct osc_create_async_args *args = ptlrpc_req_async_args(req);
  84         struct osc_creator *oscc = args->rq_oscc;
  85         struct ost_body *body = NULL;
  86         struct ptlrpc_request *fake_req, *pos;
  87         ENTRY;
  88
  89         if (req->rq_repmsg) {
  90                 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
  91                 if (body == NULL && rc == 0)
  92                         rc = -EPROTO;
  93         }
  94
  95         LASSERT(oscc && (oscc->oscc_obd != LP_POISON));
  96
  97         cfs_spin_lock(&oscc->oscc_lock);
  98         oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
  99         switch (rc) {
 100         case 0: {
 101                 if (body) {
 102                         int diff =ostid_id(&body->oa.o_oi)- oscc->oscc_last_id;
 103
 104                         /* oscc_internal_create() stores the original value of
 105                          * grow_count in osc_create_async_args::rq_grow_count.
 106                          * We can't compare against oscc_grow_count directly,
 107                          * because it may have been increased while the RPC
 108                          * is in flight, so we would always find ourselves
 109                          * having created fewer objects and decreasing the
 110                          * precreate request size.  b=18577 */
 111                         if (diff < args->rq_grow_count) {
 112                                 /* the OST has not managed to create all the
 113                                  * objects we asked for */
 114                                 oscc->oscc_grow_count = max(diff,
 115                                                             OST_MIN_PRECREATE);
 116                                 /* don't bump grow_count next time */
 117                                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 118                         } else {
 119                                 /* the OST is able to keep up with the work,
 120                                  * we could consider increasing grow_count
 121                                  * next time if needed */
 122                                 oscc->oscc_flags &= ~OSCC_FLAG_LOW;
 123                         }
 124                         oscc->oscc_last_id = ostid_id(&body->oa.o_oi);
 125                 }
 126                 cfs_spin_unlock(&oscc->oscc_lock);
 127                 break;
 128         }
 129         case -EROFS:
 130                 oscc->oscc_flags |= OSCC_FLAG_RDONLY;
 131         case -ENOSPC:
 132         case -EFBIG:
 133                 if (rc != -EROFS) {
 134                         oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 135                         if (body && rc == -ENOSPC) {
 136                                 oscc->oscc_last_id = body->oa.o_id;
 137                                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
 138
 139                                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
 140                                     (body->oa.o_flags & OBD_FL_NOSPC_BLK))
 141                                         oscc->oscc_flags |= OSCC_FLAG_NOSPC_BLK;
 142                                 else
 143                                         rc = 0;
 144                         }
 145                 }
 146                 cfs_spin_unlock(&oscc->oscc_lock);
 147                 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
 148                 break;
 149         case -EIO: {
 150                 /* filter always set body->oa.o_id as the last_id
 151                  * of filter (see filter_handle_precreate for detail)*/
 152                 if (body && body->oa.o_id > oscc->oscc_last_id)
 153                         oscc->oscc_last_id = body->oa.o_id;
 154                 cfs_spin_unlock(&oscc->oscc_lock);
 155                 break;
 156         }
 157         case -EINTR:
 158         case -EWOULDBLOCK: {
 159                 /* aka EAGAIN we should not delay create if import failed -
 160                  * this avoid client stick in create and avoid race with
 161                  * delorphan */
 162                 /* EINTR say - old create request is killed due mds<>ost
 163                  * eviction - OSCC_FLAG_RECOVERING can already set due
 164                  * IMP_DISCONN event */
 165                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 166                 /* oscc->oscc_grow_count = OST_MIN_PRECREATE; */
 167                 cfs_spin_unlock(&oscc->oscc_lock);
 168                 break;
 169         }
 170         default: {
 171                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 172                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
 173                 cfs_spin_unlock(&oscc->oscc_lock);
 174                 DEBUG_REQ(D_ERROR, req,
 175                           "Unknown rc %d from async create: failing oscc", rc);
 176                 ptlrpc_fail_import(req->rq_import,
 177                                    lustre_msg_get_conn_cnt(req->rq_reqmsg));
 178         }
 179         }
 180
 181         CDEBUG(D_HA, "preallocated through id "LPU64" (next to use "LPU64")\n",
 182                oscc->oscc_last_id, oscc->oscc_next_id);
 183
 184         cfs_spin_lock(&oscc->oscc_lock);
 185         cfs_list_for_each_entry_safe(fake_req, pos,
 186                                      &oscc->oscc_wait_create_list, rq_list) {
 187                 if (handle_async_create(fake_req, rc)  == -EAGAIN) {
 188                         oscc_internal_create(oscc);
 189                         /* sending request should be never fail because
 190                          * osc use preallocated requests pool */
 191                         GOTO(exit_wakeup, rc);
 192                 }
 193         }
 194         cfs_spin_unlock(&oscc->oscc_lock);
 195
 196 exit_wakeup:
 197         cfs_waitq_signal(&oscc->oscc_waitq);
 198         RETURN(rc);
 199 }
 200
 201 static int oscc_internal_create(struct osc_creator *oscc)
 202 {
 203         struct osc_create_async_args *args;
 204         struct ptlrpc_request *request;
 205         struct ost_body *body;
 206         ENTRY;
 207
 208         LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
 209
 210         /* Do not check for a degraded OST here - bug21563/bug18539 */
 211         if (oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 212                 cfs_spin_unlock(&oscc->oscc_lock);
 213                 RETURN(0);
 214         }
 215
 216         /* we need check it before OSCC_FLAG_CREATING - because need
 217          * see lower number of precreate objects */
 218         if (oscc->oscc_grow_count < oscc->oscc_max_grow_count &&
 219             ((oscc->oscc_flags & OSCC_FLAG_LOW) == 0) &&
 220             (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
 221                    (oscc->oscc_grow_count / 4 + 1)) {
 222                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 223                 oscc->oscc_grow_count *= 2;
 224         }
 225
 226         if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
 227                 cfs_spin_unlock(&oscc->oscc_lock);
 228                 RETURN(0);
 229         }
 230
 231         if (oscc->oscc_grow_count > oscc->oscc_max_grow_count / 2)
 232                 oscc->oscc_grow_count = oscc->oscc_max_grow_count / 2;
 233
 234         oscc->oscc_flags |= OSCC_FLAG_CREATING;
 235         cfs_spin_unlock(&oscc->oscc_lock);
 236
 237         request = ptlrpc_request_alloc_pack(oscc->oscc_obd->u.cli.cl_import,
 238                                             &RQF_OST_CREATE,
 239                                             LUSTRE_OST_VERSION, OST_CREATE);
 240         if (request == NULL) {
 241                 cfs_spin_lock(&oscc->oscc_lock);
 242                 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
 243                 cfs_spin_unlock(&oscc->oscc_lock);
 244                 RETURN(-ENOMEM);
 245         }
 246
 247         request->rq_request_portal = OST_CREATE_PORTAL;
 248         ptlrpc_at_set_req_timeout(request);
 249         body = req_capsule_client_get(&request->rq_pill, &RMF_OST_BODY);
 250         args = ptlrpc_req_async_args(request);
 251         args->rq_oscc = oscc;
 252
 253         cfs_spin_lock(&oscc->oscc_lock);
 254         args->rq_grow_count = oscc->oscc_grow_count;
 255
 256         if (likely(fid_seq_is_mdt(oscc->oscc_oa.o_seq))) {
 257                 body->oa.o_oi.oi_seq = oscc->oscc_oa.o_seq;
 258                 body->oa.o_oi.oi_id  = oscc->oscc_last_id +
 259                                        oscc->oscc_grow_count;
 260         } else {
 261                 /*Just warning here currently, since not sure how fid-on-ost
 262                  *will be implemented here */
 263                 CWARN("o_seq: "LPU64" is not indicate any MDTs.\n",
 264                        oscc->oscc_oa.o_seq);
 265         }
 266         cfs_spin_unlock(&oscc->oscc_lock);
 267
 268         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 269         CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
 270                body->oa.o_id, oscc->oscc_last_id);
 271
 272         /* we should not resend create request - anyway we will have delorphan
 273          * and kill these objects */
 274         request->rq_no_delay = request->rq_no_resend = 1;
 275         ptlrpc_request_set_replen(request);
 276
 277         request->rq_interpret_reply = osc_interpret_create;
 278         ptlrpcd_add_req(request, PSCOPE_OTHER);
 279
 280         RETURN(0);
 281 }
 282
 283 static int oscc_has_objects_nolock(struct osc_creator *oscc, int count)
 284 {
 285         return ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 286 }
 287
 288
 289 static int oscc_has_objects(struct osc_creator *oscc, int count)
 290 {
 291         int have_objs;
 292
 293         cfs_spin_lock(&oscc->oscc_lock);
 294         have_objs = oscc_has_objects_nolock(oscc, count);
 295         cfs_spin_unlock(&oscc->oscc_lock);
 296
 297         return have_objs;
 298 }
 299
 300 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 301 {
 302         int have_objs;
 303         int ost_unusable;
 304
 305         ost_unusable = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 306
 307         cfs_spin_lock(&oscc->oscc_lock);
 308         ost_unusable |= (OSCC_FLAG_NOSPC | OSCC_FLAG_RDONLY |
 309                          OSCC_FLAG_EXITING) & oscc->oscc_flags;
 310         have_objs = oscc_has_objects_nolock(oscc, count);
 311
 312         if (!ost_unusable && !have_objs)
 313                 /* they release lock himself */
 314                 have_objs = oscc_internal_create(oscc);
 315         else
 316                 cfs_spin_unlock(&oscc->oscc_lock);
 317
 318         return have_objs || ost_unusable;
 319 }
 320
 321 static int oscc_precreate(struct osc_creator *oscc)
 322 {
 323         struct l_wait_info lwi;
 324         int rc = 0;
 325         ENTRY;
 326
 327         if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
 328                 RETURN(0);
 329
 330         /* we should be not block forever - because client's create rpc can
 331          * stick in mds for long time and forbid client reconnect */
 332         lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(osc_create_timeout)),
 333                           NULL, NULL);
 334
 335         rc = l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 336         RETURN(rc);
 337 }
 338
 339 static int oscc_in_sync(struct osc_creator *oscc)
 340 {
 341         int sync;
 342
 343         cfs_spin_lock(&oscc->oscc_lock);
 344         sync = oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS;
 345         cfs_spin_unlock(&oscc->oscc_lock);
 346
 347         return sync;
 348 }
 349
 350 /* decide if the OST has remaining object, return value :
 351         0 : the OST has remaining objects, may or may not send precreation RPC.
 352         1 : the OST has no remaining object, and the sent precreation RPC
 353             has not been completed yet.
 354         2 : the OST has no remaining object, and will not get any for
 355             a potentially very long time
 356      1000 : unusable
 357  */
 358 int osc_precreate(struct obd_export *exp)
 359 {
 360         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 361         struct obd_import *imp = exp->exp_imp_reverse;
 362         int rc;
 363         ENTRY;
 364
 365         LASSERT(oscc != NULL);
 366         if (imp != NULL && imp->imp_deactive)
 367                 GOTO(out_nolock, rc = 1000);
 368
 369         /* Handle critical states first */
 370         cfs_spin_lock(&oscc->oscc_lock);
 371         if (oscc->oscc_flags & OSCC_FLAG_NOSPC_BLK ||
 372             oscc->oscc_flags & OSCC_FLAG_RDONLY ||
 373             oscc->oscc_flags & OSCC_FLAG_EXITING)
 374                 GOTO(out, rc = 1000);
 375
 376         if ((oscc->oscc_flags & OSCC_FLAG_RECOVERING) ||
 377             (oscc->oscc_flags & OSCC_FLAG_DEGRADED))
 378                 GOTO(out, rc = 2);
 379
 380         if (oscc_has_objects_nolock(oscc, oscc->oscc_grow_count / 2))
 381                 GOTO(out, rc = 0);
 382
 383         /* Return 0, if we have at least one object - bug 22884 */
 384         rc = oscc_has_objects_nolock(oscc, 1) ? 0 : 1;
 385
 386         if (oscc->oscc_flags & OSCC_FLAG_NOSPC)
 387                 GOTO(out, (rc == 0) ? 0 : 1000);
 388
 389         /* Do not check for OSCC_FLAG_CREATING flag here, let
 390          * osc_precreate() call oscc_internal_create() and
 391          * adjust oscc_grow_count bug21563 */
 392         if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS)
 393                 GOTO(out, rc);
 394
 395         if (oscc_internal_create(oscc))
 396                 GOTO(out_nolock, rc = 1000);
 397
 398         RETURN(rc);
 399 out:
 400         cfs_spin_unlock(&oscc->oscc_lock);
 401 out_nolock:
 402         return rc;
 403 }
 404
 405 static int handle_async_create(struct ptlrpc_request *req, int rc)
 406 {
 407         struct osc_create_async_args *args = ptlrpc_req_async_args(req);
 408         struct osc_creator    *oscc = args->rq_oscc;
 409         struct lov_stripe_md  *lsm  = args->rq_lsm;
 410         struct obd_info       *oinfo = args->rq_oinfo;
 411         struct obdo           *oa = oinfo->oi_oa;
 412
 413         LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
 414
 415         if(rc)
 416                 GOTO(out_wake, rc);
 417
 418         /* Handle the critical type errors first.
 419          * Should we also test cl_import state as well ? */
 420         if (oscc->oscc_flags & OSCC_FLAG_EXITING)
 421                 GOTO(out_wake, rc = -EIO);
 422
 423         if (oscc->oscc_flags & OSCC_FLAG_NOSPC_BLK)
 424                 GOTO(out_wake, rc = -ENOSPC);
 425
 426         if (oscc->oscc_flags & OSCC_FLAG_RDONLY)
 427                 GOTO(out_wake, rc = -EROFS);
 428
 429         /* should be try wait until recovery finished */
 430         if((oscc->oscc_flags & OSCC_FLAG_RECOVERING) ||
 431            (oscc->oscc_flags & OSCC_FLAG_DEGRADED))
 432                 RETURN(-EAGAIN);
 433
 434         if (oscc_has_objects_nolock(oscc, 1)) {
 435                 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 436                 oa->o_id = oscc->oscc_next_id;
 437                 lsm->lsm_object_id = oscc->oscc_next_id;
 438                 oscc->oscc_next_id++;
 439
 440                 CDEBUG(D_RPCTRACE, " set oscc_next_id = "LPU64"\n",
 441                        oscc->oscc_next_id);
 442                 GOTO(out_wake, rc = 0);
 443         }
 444
 445         /* we don't have objects now - continue wait */
 446         RETURN(-EAGAIN);
 447
 448 out_wake:
 449
 450         rc = oinfo->oi_cb_up(oinfo, rc);
 451         ptlrpc_fakereq_finished(req);
 452
 453         RETURN(rc);
 454 }
 455
 456 static int async_create_interpret(const struct lu_env *env,
 457                                   struct ptlrpc_request *req, void *data,
 458                                   int rc)
 459 {
 460         struct osc_create_async_args *args = ptlrpc_req_async_args(req);
 461         struct osc_creator    *oscc = args->rq_oscc;
 462         int ret;
 463
 464         cfs_spin_lock(&oscc->oscc_lock);
 465         ret = handle_async_create(req, rc);
 466         cfs_spin_unlock(&oscc->oscc_lock);
 467
 468         return ret;
 469 }
 470
 471 int osc_create_async(struct obd_export *exp, struct obd_info *oinfo,
 472                      struct lov_stripe_md **ea, struct obd_trans_info *oti)
 473 {
 474         int rc;
 475         struct ptlrpc_request *fake_req;
 476         struct osc_create_async_args *args;
 477         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 478         struct obdo *oa = oinfo->oi_oa;
 479         ENTRY;
 480
 481         if ((oa->o_valid & OBD_MD_FLGROUP) && !fid_seq_is_mdt(oa->o_seq)) {
 482                 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
 483                 rc = oinfo->oi_cb_up(oinfo, rc);
 484                 RETURN(rc);
 485         }
 486
 487         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 488             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 489                 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
 490                 rc = oinfo->oi_cb_up(oinfo, rc);
 491                 RETURN(rc);
 492         }
 493
 494         LASSERT((*ea) != NULL);
 495
 496         fake_req = ptlrpc_prep_fakereq(oscc->oscc_obd->u.cli.cl_import,
 497                                        osc_create_timeout,
 498                                        async_create_interpret);
 499         if (fake_req == NULL) {
 500                 rc = oinfo->oi_cb_up(oinfo, -ENOMEM);
 501                 RETURN(-ENOMEM);
 502         }
 503
 504         args = ptlrpc_req_async_args(fake_req);
 505         CLASSERT(sizeof(*args) <= sizeof(fake_req->rq_async_args));
 506
 507         args->rq_oscc  = oscc;
 508         args->rq_lsm   = *ea;
 509         args->rq_oinfo = oinfo;
 510
 511         cfs_spin_lock(&oscc->oscc_lock);
 512         /* try fast path */
 513         rc = handle_async_create(fake_req, 0);
 514         if (rc == -EAGAIN) {
 515                 int is_add;
 516                 /* we not have objects - try wait */
 517                 is_add = ptlrpcd_add_req(fake_req, PSCOPE_OTHER);
 518                 if (!is_add)
 519                         cfs_list_add(&fake_req->rq_list,
 520                                      &oscc->oscc_wait_create_list);
 521                 else
 522                         rc = is_add;
 523         }
 524         cfs_spin_unlock(&oscc->oscc_lock);
 525
 526         if (rc != -EAGAIN)
 527                 /* need free request if was error hit or
 528                  * objects already allocated */
 529                 ptlrpc_req_finished(fake_req);
 530         else
 531                 /* EAGAIN mean - request is delayed */
 532                 rc = 0;
 533
 534         RETURN(rc);
 535 }
 536
 537 int osc_create(struct obd_export *exp, struct obdo *oa,
 538                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 539 {
 540         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 541         struct obd_import  *imp  = exp->exp_obd->u.cli.cl_import;
 542         struct lov_stripe_md *lsm;
 543         int del_orphan = 0, rc = 0;
 544         ENTRY;
 545
 546         LASSERT(oa);
 547         LASSERT(ea);
 548         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 549
 550         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 551             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 552                 RETURN(osc_real_create(exp, oa, ea, oti));
 553         }
 554
 555         if (!fid_seq_is_mdt(oa->o_seq))
 556                 RETURN(osc_real_create(exp, oa, ea, oti));
 557
 558         /* this is the special case where create removes orphans */
 559         if (oa->o_valid & OBD_MD_FLFLAGS &&
 560             oa->o_flags == OBD_FL_DELORPHAN) {
 561                 cfs_spin_lock(&oscc->oscc_lock);
 562                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 563                         cfs_spin_unlock(&oscc->oscc_lock);
 564                         RETURN(-EBUSY);
 565                 }
 566                 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 567                         cfs_spin_unlock(&oscc->oscc_lock);
 568                         RETURN(0);
 569                 }
 570
 571                 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
 572                 /* seting flag LOW we prevent extra grow precreate size
 573                  * and enforce use last assigned size */
 574                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 575                 cfs_spin_unlock(&oscc->oscc_lock);
 576                 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
 577                        oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
 578
 579                 del_orphan = 1;
 580
 581                 /* delete from next_id on up */
 582                 oa->o_valid |= OBD_MD_FLID;
 583                 oa->o_id = oscc->oscc_next_id - 1;
 584
 585                 rc = osc_real_create(exp, oa, ea, NULL);
 586
 587                 cfs_spin_lock(&oscc->oscc_lock);
 588                 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
 589                 if (rc == 0 || rc == -ENOSPC) {
 590                         struct obd_connect_data *ocd;
 591
 592                         if (rc == -ENOSPC) {
 593                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 594                                 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 595                                     (oa->o_flags & OBD_FL_NOSPC_BLK))
 596                                         oscc->oscc_flags |= OSCC_FLAG_NOSPC_BLK;
 597                         }
 598                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 599
 600                         oscc->oscc_last_id = oa->o_id;
 601                         ocd = &imp->imp_connect_data;
 602                         if (ocd->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN) {
 603                                 /*
 604                                  * The OST reports back in oa->o_id from where
 605                                  * we should restart in order to skip orphan
 606                                  * objects
 607                                  */
 608                                 CDEBUG(D_HA, "%s: Skip orphan set, reset last "
 609                                        "objid\n", oscc->oscc_obd->obd_name);
 610                                 oscc->oscc_next_id = oa->o_id + 1;
 611                         }
 612
 613                         /* sanity check for next objid. see bug 17025 */
 614                         LASSERT(oscc->oscc_next_id == oa->o_id + 1);
 615
 616                         CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
 617                                LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
 618                                oscc->oscc_last_id, rc);
 619                 } else {
 620                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 621                                oscc->oscc_obd->obd_name, rc);
 622                 }
 623
 624                 cfs_waitq_signal(&oscc->oscc_waitq);
 625                 cfs_spin_unlock(&oscc->oscc_lock);
 626
 627                 if (rc < 0)
 628                         RETURN(rc);
 629         }
 630
 631         lsm = *ea;
 632         if (lsm == NULL) {
 633                 rc = obd_alloc_memmd(exp, &lsm);
 634                 if (rc < 0)
 635                         RETURN(rc);
 636         }
 637
 638         while (1) {
 639                 if (oscc_in_sync(oscc))
 640                         CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
 641                                oscc->oscc_obd->obd_name);
 642
 643                 rc = oscc_precreate(oscc);
 644                 if (rc)
 645                         CDEBUG(D_HA,"%s: error create %d\n",
 646                                oscc->oscc_obd->obd_name, rc);
 647
 648                 cfs_spin_lock(&oscc->oscc_lock);
 649
 650                 /* wakeup but recovery did not finished */
 651                 if ((oscc->oscc_obd->u.cli.cl_import->imp_invalid) ||
 652                     (oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 653                         rc = -EIO;
 654                         cfs_spin_unlock(&oscc->oscc_lock);
 655                         break;
 656                 }
 657
 658                 if (oscc->oscc_flags & OSCC_FLAG_NOSPC_BLK) {
 659                         rc = -ENOSPC;
 660                         cfs_spin_unlock(&oscc->oscc_lock);
 661                         break;
 662                 }
 663
 664                 if (oscc->oscc_flags & OSCC_FLAG_RDONLY) {
 665                         rc = -EROFS;
 666                         cfs_spin_unlock(&oscc->oscc_lock);
 667                         break;
 668                 }
 669
 670                 // Should we report -EIO error ?
 671                 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
 672                         cfs_spin_unlock(&oscc->oscc_lock);
 673                         break;
 674                 }
 675
 676                 /**
 677                  * If this is DELORPHAN process, no need create object here,
 678                  * otherwise this will create a gap of object id, and MDS
 679                  * might create some orphan log (mds_lov_update_objids), then
 680                  * remove objects wrongly on OST. Bug 21379.
 681                  */
 682                 if (oa->o_valid & OBD_MD_FLFLAGS &&
 683                         oa->o_flags == OBD_FL_DELORPHAN) {
 684                         cfs_spin_unlock(&oscc->oscc_lock);
 685                         break;
 686                 }
 687
 688                 if (oscc_has_objects_nolock(oscc, 1)) {
 689                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 690                         oa->o_id = oscc->oscc_next_id;
 691                         lsm->lsm_object_id = oscc->oscc_next_id;
 692                         *ea = lsm;
 693                         oscc->oscc_next_id++;
 694                         cfs_spin_unlock(&oscc->oscc_lock);
 695
 696                         CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
 697                                exp->exp_obd->obd_name, oscc->oscc_next_id);
 698                         break;
 699                 }
 700
 701                 if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 702                         rc = -ENOSPC;
 703                         cfs_spin_unlock(&oscc->oscc_lock);
 704                         break;
 705                 }
 706
 707                 cfs_spin_unlock(&oscc->oscc_lock);
 708         }
 709
 710         if (rc == 0) {
 711                 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
 712                        obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
 713         } else {
 714                 if (*ea == NULL)
 715                         obd_free_memmd(exp, &lsm);
 716                 if (del_orphan != 0 && rc != -EIO)
 717                         /* Ignore non-IO precreate error for clear orphan */
 718                         rc = 0;
 719         }
 720         RETURN(rc);
 721 }
 722
 723 void oscc_init(struct obd_device *obd)
 724 {
 725         struct osc_creator *oscc;
 726
 727         if (obd == NULL)
 728                 return;
 729
 730         oscc = &obd->u.cli.cl_oscc;
 731
 732         memset(oscc, 0, sizeof(*oscc));
 733
 734         cfs_waitq_init(&oscc->oscc_waitq);
 735         cfs_spin_lock_init(&oscc->oscc_lock);
 736         oscc->oscc_obd = obd;
 737         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 738         oscc->oscc_max_grow_count = OST_MAX_PRECREATE;
 739
 740         oscc->oscc_next_id = 2;
 741         oscc->oscc_last_id = 1;
 742         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 743
 744         CFS_INIT_LIST_HEAD(&oscc->oscc_wait_create_list);
 745
 746         /* XXX the export handle should give the oscc the last object */
 747         /* oed->oed_oscc.oscc_last_id = exph->....; */
 748 }
 749
 750 void oscc_fini(struct obd_device *obd)
 751 {
 752         struct osc_creator *oscc = &obd->u.cli.cl_oscc;
 753         ENTRY;
 754
 755
 756         cfs_spin_lock(&oscc->oscc_lock);
 757         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 758         oscc->oscc_flags |= OSCC_FLAG_EXITING;
 759         cfs_spin_unlock(&oscc->oscc_lock);
 760 }