lustre/osp/osp_dev.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, Intel Corporation.
  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/osp/osp_dev.c
  37  *
  38  * Lustre OST Proxy Device
  39  *
  40  * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
  41  * Author: Mikhail Pershin <mike.pershin@intel.com>
  42  * Author: Di Wang <di.wang@intel.com>
  43  */
  44
  45 #ifndef EXPORT_SYMTAB
  46 # define EXPORT_SYMTAB
  47 #endif
  48 #define DEBUG_SUBSYSTEM S_MDS
  49
  50 #include <obd_class.h>
  51 #include <lustre_param.h>
  52 #include <lustre_log.h>
  53 #include <lustre_mdc.h>
  54
  55 #include "osp_internal.h"
  56
  57 /* Slab for OSP object allocation */
  58 cfs_mem_cache_t *osp_object_kmem;
  59
  60 static struct lu_kmem_descr osp_caches[] = {
  61         {
  62                 .ckd_cache = &osp_object_kmem,
  63                 .ckd_name  = "osp_obj",
  64                 .ckd_size  = sizeof(struct osp_object)
  65         },
  66         {
  67                 .ckd_cache = NULL
  68         }
  69 };
  70
  71 struct lu_object *osp_object_alloc(const struct lu_env *env,
  72                                    const struct lu_object_header *hdr,
  73                                    struct lu_device *d)
  74 {
  75         struct lu_object_header *h;
  76         struct osp_object       *o;
  77         struct lu_object        *l;
  78
  79         LASSERT(hdr == NULL);
  80
  81         OBD_SLAB_ALLOC_PTR_GFP(o, osp_object_kmem, CFS_ALLOC_IO);
  82         if (o != NULL) {
  83                 l = &o->opo_obj.do_lu;
  84                 h = &o->opo_header;
  85
  86                 lu_object_header_init(h);
  87                 dt_object_init(&o->opo_obj, h, d);
  88                 lu_object_add_top(h, l);
  89
  90                 l->lo_ops = &osp_lu_obj_ops;
  91
  92                 return l;
  93         } else {
  94                 return NULL;
  95         }
  96 }
  97
  98 static struct dt_object
  99 *osp_find_or_create(const struct lu_env *env, struct osp_device *osp,
 100                     struct lu_attr *attr, __u32 reg_id)
 101 {
 102         struct osp_thread_info *osi = osp_env_info(env);
 103         struct dt_object_format dof = { 0 };
 104         struct dt_object       *dto;
 105         int                  rc;
 106         ENTRY;
 107
 108         lu_local_obj_fid(&osi->osi_fid, reg_id);
 109         attr->la_valid = LA_MODE;
 110         attr->la_mode = S_IFREG | 0644;
 111         dof.dof_type = DFT_REGULAR;
 112         dto = dt_find_or_create(env, osp->opd_storage, &osi->osi_fid,
 113                                 &dof, attr);
 114         if (IS_ERR(dto))
 115                 RETURN(dto);
 116
 117         rc = dt_attr_get(env, dto, attr, NULL);
 118         if (rc) {
 119                 CERROR("%s: can't be initialized: rc = %d\n",
 120                        osp->opd_obd->obd_name, rc);
 121                 lu_object_put(env, &dto->do_lu);
 122                 RETURN(ERR_PTR(rc));
 123         }
 124         RETURN(dto);
 125 }
 126
 127 static int osp_write_local_file(const struct lu_env *env,
 128                                 struct osp_device *osp,
 129                                 struct dt_object *dt_obj,
 130                                 struct lu_buf *buf,
 131                                 loff_t offset)
 132 {
 133         struct thandle *th;
 134         int rc;
 135
 136         th = dt_trans_create(env, osp->opd_storage);
 137         if (IS_ERR(th))
 138                 RETURN(PTR_ERR(th));
 139
 140         rc = dt_declare_record_write(env, dt_obj, buf->lb_len, offset, th);
 141         if (rc)
 142                 GOTO(out, rc);
 143         rc = dt_trans_start_local(env, osp->opd_storage, th);
 144         if (rc)
 145                 GOTO(out, rc);
 146
 147         rc = dt_record_write(env, dt_obj, buf, &offset, th);
 148 out:
 149         dt_trans_stop(env, osp->opd_storage, th);
 150         RETURN(rc);
 151 }
 152
 153 static int osp_init_last_objid(const struct lu_env *env, struct osp_device *osp)
 154 {
 155         struct osp_thread_info  *osi = osp_env_info(env);
 156         struct lu_fid           *fid = &osp->opd_last_used_fid;
 157         struct dt_object        *dto;
 158         int                     rc;
 159         ENTRY;
 160
 161         dto = osp_find_or_create(env, osp, &osi->osi_attr, MDD_LOV_OBJ_OID);
 162         if (IS_ERR(dto))
 163                 RETURN(PTR_ERR(dto));
 164         /* object will be released in device cleanup path */
 165         if (osi->osi_attr.la_size >=
 166             sizeof(osi->osi_id) * (osp->opd_index + 1)) {
 167                 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
 168                                    osp->opd_index);
 169                 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
 170                 if (rc != 0)
 171                         GOTO(out, rc);
 172         } else {
 173                 fid->f_oid = 0;
 174                 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
 175                                    osp->opd_index);
 176                 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
 177                                           osi->osi_off);
 178         }
 179         osp->opd_last_used_oid_file = dto;
 180         RETURN(0);
 181 out:
 182         /* object will be released in device cleanup path */
 183         CERROR("%s: can't initialize lov_objid: rc = %d\n",
 184                osp->opd_obd->obd_name, rc);
 185         lu_object_put(env, &dto->do_lu);
 186         osp->opd_last_used_oid_file = NULL;
 187         RETURN(rc);
 188 }
 189
 190 static int osp_init_last_seq(const struct lu_env *env, struct osp_device *osp)
 191 {
 192         struct osp_thread_info  *osi = osp_env_info(env);
 193         struct lu_fid           *fid = &osp->opd_last_used_fid;
 194         struct dt_object        *dto;
 195         int                     rc;
 196         ENTRY;
 197
 198         dto = osp_find_or_create(env, osp, &osi->osi_attr, MDD_LOV_OBJ_OSEQ);
 199         if (IS_ERR(dto))
 200                 RETURN(PTR_ERR(dto));
 201
 202         /* object will be released in device cleanup path */
 203         if (osi->osi_attr.la_size >=
 204             sizeof(osi->osi_id) * (osp->opd_index + 1)) {
 205                 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
 206                                    osp->opd_index);
 207                 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
 208                 if (rc != 0)
 209                         GOTO(out, rc);
 210         } else {
 211                 fid->f_seq = 0;
 212                 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
 213                                     osp->opd_index);
 214                 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
 215                                           osi->osi_off);
 216         }
 217         osp->opd_last_used_seq_file = dto;
 218         RETURN(0);
 219 out:
 220         /* object will be released in device cleanup path */
 221         CERROR("%s: can't initialize lov_seq: rc = %d\n",
 222                osp->opd_obd->obd_name, rc);
 223         lu_object_put(env, &dto->do_lu);
 224         osp->opd_last_used_seq_file = NULL;
 225         RETURN(rc);
 226 }
 227
 228 static int osp_last_used_init(const struct lu_env *env, struct osp_device *osp)
 229 {
 230         struct osp_thread_info *osi = osp_env_info(env);
 231         int                  rc;
 232         ENTRY;
 233
 234         fid_zero(&osp->opd_last_used_fid);
 235         rc = osp_init_last_objid(env, osp);
 236         if (rc < 0) {
 237                 CERROR("%s: Can not get ids %d from old objid!\n",
 238                        osp->opd_obd->obd_name, rc);
 239                 RETURN(rc);
 240         }
 241
 242         rc = osp_init_last_seq(env, osp);
 243         if (rc < 0) {
 244                 CERROR("%s: Can not get ids %d from old objid!\n",
 245                        osp->opd_obd->obd_name, rc);
 246                 GOTO(out, rc);
 247         }
 248
 249         if (fid_oid(&osp->opd_last_used_fid) != 0 &&
 250             fid_seq(&osp->opd_last_used_fid) == 0) {
 251                 /* Just upgrade from the old version,
 252                  * set the seq to be IDIF */
 253                 osp->opd_last_used_fid.f_seq =
 254                    fid_idif_seq(fid_oid(&osp->opd_last_used_fid),
 255                                 osp->opd_index);
 256                 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off,
 257                                     &osp->opd_last_used_fid.f_seq,
 258                                     osp->opd_index);
 259                 rc = osp_write_local_file(env, osp, osp->opd_last_used_seq_file,
 260                                           &osi->osi_lb, osi->osi_off);
 261                 if (rc) {
 262                         CERROR("%s : Can not write seq file: rc = %d\n",
 263                                osp->opd_obd->obd_name, rc);
 264                         GOTO(out, rc);
 265                 }
 266         }
 267
 268         if (!fid_is_zero(&osp->opd_last_used_fid) &&
 269                  !fid_is_sane(&osp->opd_last_used_fid)) {
 270                 CERROR("%s: Got invalid FID "DFID"\n", osp->opd_obd->obd_name,
 271                         PFID(&osp->opd_last_used_fid));
 272                 GOTO(out, rc = -EINVAL);
 273         }
 274
 275         CDEBUG(D_INFO, "%s: Init last used fid "DFID"\n",
 276                osp->opd_obd->obd_name, PFID(&osp->opd_last_used_fid));
 277 out:
 278         if (rc != 0) {
 279                 if (osp->opd_last_used_oid_file != NULL) {
 280                         lu_object_put(env, &osp->opd_last_used_oid_file->do_lu);
 281                         osp->opd_last_used_oid_file = NULL;
 282                 }
 283                 if (osp->opd_last_used_seq_file != NULL) {
 284                         lu_object_put(env, &osp->opd_last_used_seq_file->do_lu);
 285                         osp->opd_last_used_seq_file = NULL;
 286                 }
 287         }
 288
 289         RETURN(rc);
 290 }
 291
 292 static void osp_last_used_fini(const struct lu_env *env, struct osp_device *d)
 293 {
 294         /* release last_used file */
 295         if (d->opd_last_used_oid_file != NULL) {
 296                 lu_object_put(env, &d->opd_last_used_oid_file->do_lu);
 297                 d->opd_last_used_oid_file = NULL;
 298         }
 299
 300         if (d->opd_last_used_seq_file != NULL) {
 301                 lu_object_put(env, &d->opd_last_used_seq_file->do_lu);
 302                 d->opd_last_used_seq_file = NULL;
 303         }
 304 }
 305
 306 static int osp_disconnect(struct osp_device *d)
 307 {
 308         struct obd_import *imp;
 309         int rc = 0;
 310
 311         imp = d->opd_obd->u.cli.cl_import;
 312
 313         /* Mark import deactivated now, so we don't try to reconnect if any
 314          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 315          * fully deactivate the import, or that would drop all requests. */
 316         LASSERT(imp != NULL);
 317         spin_lock(&imp->imp_lock);
 318         imp->imp_deactive = 1;
 319         spin_unlock(&imp->imp_lock);
 320
 321         ptlrpc_deactivate_import(imp);
 322
 323         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 324          * delete it regardless.  (It's safe to delete an import that was
 325          * never added.) */
 326         (void)ptlrpc_pinger_del_import(imp);
 327
 328         rc = ptlrpc_disconnect_import(imp, 0);
 329         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
 330                 rc = 0;
 331         if (rc)
 332                 CERROR("%s: can't disconnect: rc = %d\n",
 333                        d->opd_obd->obd_name, rc);
 334
 335         ptlrpc_invalidate_import(imp);
 336
 337         RETURN(rc);
 338 }
 339
 340 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
 341 {
 342         int                      rc = 0;
 343         ENTRY;
 344
 345         LASSERT(env);
 346         /* release last_used file */
 347         if (!d->opd_connect_mdt)
 348                 osp_last_used_fini(env, d);
 349
 350         rc = osp_disconnect(d);
 351
 352         if (!d->opd_connect_mdt) {
 353                 /* stop precreate thread */
 354                 osp_precreate_fini(d);
 355
 356                 /* stop sync thread */
 357                 osp_sync_fini(d);
 358         }
 359
 360         obd_fid_fini(d->opd_obd);
 361
 362         RETURN(rc);
 363 }
 364
 365 static int osp_process_config(const struct lu_env *env,
 366                               struct lu_device *dev, struct lustre_cfg *lcfg)
 367 {
 368         struct osp_device               *d = lu2osp_dev(dev);
 369         struct lprocfs_static_vars       lvars = { 0 };
 370         int                              rc;
 371
 372         ENTRY;
 373
 374         switch (lcfg->lcfg_command) {
 375         case LCFG_CLEANUP:
 376                 lu_dev_del_linkage(dev->ld_site, dev);
 377                 rc = osp_shutdown(env, d);
 378                 break;
 379         case LCFG_PARAM:
 380                 lprocfs_osp_init_vars(&lvars);
 381
 382                 LASSERT(d->opd_obd);
 383                 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
 384                                               lcfg, d->opd_obd);
 385                 if (rc > 0)
 386                         rc = 0;
 387                 if (rc == -ENOSYS) {
 388                         /* class_process_proc_param() haven't found matching
 389                          * parameter and returned ENOSYS so that layer(s)
 390                          * below could use that. But OSP is the bottom, so
 391                          * just ignore it */
 392                         CERROR("%s: unknown param %s\n",
 393                                (char *)lustre_cfg_string(lcfg, 0),
 394                                (char *)lustre_cfg_string(lcfg, 1));
 395                         rc = 0;
 396                 }
 397                 break;
 398         default:
 399                 CERROR("%s: unknown command %u\n",
 400                        (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
 401                 rc = 0;
 402                 break;
 403         }
 404
 405         RETURN(rc);
 406 }
 407
 408 static int osp_recovery_complete(const struct lu_env *env,
 409                                  struct lu_device *dev)
 410 {
 411         struct osp_device       *osp = lu2osp_dev(dev);
 412         int                      rc = 0;
 413
 414         ENTRY;
 415         osp->opd_recovery_completed = 1;
 416         if (!osp->opd_connect_mdt)
 417                 cfs_waitq_signal(&osp->opd_pre_waitq);
 418         RETURN(rc);
 419 }
 420
 421 const struct lu_device_operations osp_lu_ops = {
 422         .ldo_object_alloc       = osp_object_alloc,
 423         .ldo_process_config     = osp_process_config,
 424         .ldo_recovery_complete  = osp_recovery_complete,
 425 };
 426
 427 /**
 428  * provides with statfs from corresponded OST
 429  *
 430  */
 431 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
 432                       struct obd_statfs *sfs)
 433 {
 434         struct osp_device *d = dt2osp_dev(dev);
 435
 436         ENTRY;
 437
 438         if (unlikely(d->opd_imp_active == 0))
 439                 RETURN(-ENOTCONN);
 440
 441         /* return recently updated data */
 442         *sfs = d->opd_statfs;
 443
 444         /*
 445          * layer above osp (usually lod) can use ffree to estimate
 446          * how many objects are available for immediate creation
 447          */
 448
 449         spin_lock(&d->opd_pre_lock);
 450         LASSERTF(fid_seq(&d->opd_pre_last_created_fid) ==
 451                  fid_seq(&d->opd_pre_used_fid),
 452                  "last_created "DFID", next_fid "DFID"\n",
 453                  PFID(&d->opd_pre_last_created_fid),
 454                  PFID(&d->opd_pre_used_fid));
 455         sfs->os_fprecreated = fid_oid(&d->opd_pre_last_created_fid) -
 456                               fid_oid(&d->opd_pre_used_fid);
 457         sfs->os_fprecreated -= d->opd_pre_reserved;
 458         spin_unlock(&d->opd_pre_lock);
 459
 460         LASSERT(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2);
 461
 462         CDEBUG(D_OTHER, "%s: "LPU64" blocks, "LPU64" free, "LPU64" avail, "
 463                LPU64" files, "LPU64" free files\n", d->opd_obd->obd_name,
 464                sfs->os_blocks, sfs->os_bfree, sfs->os_bavail,
 465                sfs->os_files, sfs->os_ffree);
 466         RETURN(0);
 467 }
 468
 469 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
 470 {
 471         ENTRY;
 472
 473         /*
 474          * XXX: wake up sync thread, command it to start flushing asap?
 475          */
 476
 477         RETURN(0);
 478 }
 479
 480 const struct dt_device_operations osp_dt_ops = {
 481         .dt_statfs      = osp_statfs,
 482         .dt_sync        = osp_sync,
 483 };
 484
 485 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *m,
 486                               const char *nextdev)
 487 {
 488         struct obd_connect_data *data = NULL;
 489         struct obd_device       *obd;
 490         int                      rc;
 491
 492         ENTRY;
 493
 494         LASSERT(m->opd_storage_exp == NULL);
 495
 496         OBD_ALLOC_PTR(data);
 497         if (data == NULL)
 498                 RETURN(-ENOMEM);
 499
 500         obd = class_name2obd(nextdev);
 501         if (obd == NULL) {
 502                 CERROR("%s: can't locate next device: %s\n",
 503                        m->opd_obd->obd_name, nextdev);
 504                 GOTO(out, rc = -ENOTCONN);
 505         }
 506
 507         rc = obd_connect(env, &m->opd_storage_exp, obd, &obd->obd_uuid, data,
 508                          NULL);
 509         if (rc) {
 510                 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
 511                        m->opd_obd->obd_name, nextdev, rc);
 512                 GOTO(out, rc);
 513         }
 514
 515         m->opd_dt_dev.dd_lu_dev.ld_site =
 516                 m->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
 517         LASSERT(m->opd_dt_dev.dd_lu_dev.ld_site);
 518         m->opd_storage = lu2dt_dev(m->opd_storage_exp->exp_obd->obd_lu_dev);
 519
 520 out:
 521         OBD_FREE_PTR(data);
 522         RETURN(rc);
 523 }
 524
 525 static int osp_init0(const struct lu_env *env, struct osp_device *m,
 526                      struct lu_device_type *ldt, struct lustre_cfg *cfg)
 527 {
 528         struct obd_device       *obd;
 529         struct obd_import       *imp;
 530         class_uuid_t            uuid;
 531         char                    *src, *tgt, *mdt, *osdname = NULL;
 532         int                     rc, idx;
 533
 534         ENTRY;
 535
 536         obd = class_name2obd(lustre_cfg_string(cfg, 0));
 537         if (obd == NULL) {
 538                 CERROR("Cannot find obd with name %s\n",
 539                        lustre_cfg_string(cfg, 0));
 540                 RETURN(-ENODEV);
 541         }
 542         m->opd_obd = obd;
 543
 544         /* There is no record in the MDT configuration for the local disk
 545          * device, so we have to extract this from elsewhere in the profile.
 546          * The only information we get at setup is from the OSC records:
 547          * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 548          * Note that 1.8 generated configs are missing the -MDTxxxx part.
 549          * We need to reconstruct the name of the underlying OSD from this:
 550          * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".  We
 551          * also need to determine the OST index from this - will be used
 552          * to calculate the offset in shared lov_objids file later */
 553
 554         src = lustre_cfg_string(cfg, 0);
 555         if (src == NULL)
 556                 RETURN(-EINVAL);
 557
 558         tgt = strrchr(src, '-');
 559         if (tgt == NULL) {
 560                 CERROR("%s: invalid target name %s\n",
 561                        m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 562                 RETURN(-EINVAL);
 563         }
 564
 565         if (strncmp(tgt, "-osc", 4) == 0) {
 566                 /* Old OSC name fsname-OSTXXXX-osc */
 567                 for (tgt--; tgt > src && *tgt != '-'; tgt--)
 568                         ;
 569                 if (tgt == src) {
 570                         CERROR("%s: invalid target name %s\n",
 571                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 572                         RETURN(-EINVAL);
 573                 }
 574
 575                 if (strncmp(tgt, "-OST", 4) != 0) {
 576                         CERROR("%s: invalid target name %s\n",
 577                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 578                         RETURN(-EINVAL);
 579                 }
 580
 581                 idx = simple_strtol(tgt + 4, &mdt, 16);
 582                 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
 583                         CERROR("%s: invalid OST index in '%s'\n",
 584                                m->opd_obd->obd_name, src);
 585                         RETURN(-EINVAL);
 586                 }
 587                 m->opd_index = idx;
 588                 idx = tgt - src;
 589         } else {
 590                 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
 591                 if (strncmp(tgt, "-MDT", 4) != 0 &&
 592                          strncmp(tgt, "-OST", 4) != 0) {
 593                         CERROR("%s: invalid target name %s\n",
 594                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 595                         RETURN(-EINVAL);
 596                 }
 597
 598                 if (tgt - src <= 12) {
 599                         CERROR("%s: invalid target name %s\n",
 600                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 601                         RETURN(-EINVAL);
 602                 }
 603
 604                 if (strncmp(tgt - 12, "-MDT", 4) == 0)
 605                         m->opd_connect_mdt = 1;
 606
 607                 idx = simple_strtol(tgt - 8, &mdt, 16);
 608                 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
 609                         CERROR("%s: invalid OST index in '%s'\n",
 610                                m->opd_obd->obd_name, src);
 611                         RETURN(-EINVAL);
 612                 }
 613
 614                 m->opd_index = idx;
 615                 idx = tgt - src - 12;
 616         }
 617         /* check the fsname length, and after this everything else will fit */
 618         if (idx > MTI_NAME_MAXLEN) {
 619                 CERROR("%s: fsname too long in '%s'\n",
 620                        m->opd_obd->obd_name, src);
 621                 RETURN(-EINVAL);
 622         }
 623
 624         OBD_ALLOC(osdname, MAX_OBD_NAME);
 625         if (osdname == NULL)
 626                 RETURN(-ENOMEM);
 627
 628         memcpy(osdname, src, idx); /* copy just the fsname part */
 629         osdname[idx] = '\0';
 630
 631         mdt = strstr(mdt, "-MDT");
 632         if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
 633                 strcat(osdname, "-MDT0000");
 634         else
 635                 strcat(osdname, mdt);
 636         strcat(osdname, "-osd");
 637         CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
 638
 639         if (m->opd_connect_mdt) {
 640                 struct client_obd *cli = &m->opd_obd->u.cli;
 641
 642                 OBD_ALLOC(cli->cl_rpc_lock, sizeof(*cli->cl_rpc_lock));
 643                 if (!cli->cl_rpc_lock)
 644                         RETURN(-ENOMEM);
 645                 osp_init_rpc_lock(cli->cl_rpc_lock);
 646         }
 647
 648         m->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
 649         m->opd_dt_dev.dd_ops = &osp_dt_ops;
 650         obd->obd_lu_dev = &m->opd_dt_dev.dd_lu_dev;
 651
 652         rc = osp_connect_to_osd(env, m, osdname);
 653         if (rc)
 654                 GOTO(out_fini, rc);
 655
 656         rc = ptlrpcd_addref();
 657         if (rc)
 658                 GOTO(out_disconnect, rc);
 659
 660         rc = client_obd_setup(obd, cfg);
 661         if (rc) {
 662                 CERROR("%s: can't setup obd: %d\n", m->opd_obd->obd_name, rc);
 663                 GOTO(out_ref, rc);
 664         }
 665
 666         osp_lprocfs_init(m);
 667
 668         if (!m->opd_connect_mdt) {
 669                 /* Initialize last id from the storage - will be
 670                  * used in orphan cleanup. */
 671                 rc = osp_last_used_init(env, m);
 672                 if (rc)
 673                         GOTO(out_proc, rc);
 674                 /* Initialize precreation thread, it handles new
 675                  * connections as well. */
 676                 rc = osp_init_precreate(m);
 677                 if (rc)
 678                         GOTO(out_last_used, rc);
 679                 /*
 680                  * Initialize synhronization mechanism taking
 681                  * care of propogating changes to OST in near
 682                  * transactional manner.
 683                  */
 684                 rc = osp_sync_init(env, m);
 685                 if (rc)
 686                         GOTO(out_precreat, rc);
 687
 688                 rc = obd_fid_init(m->opd_obd, NULL, LUSTRE_SEQ_DATA);
 689                 if (rc) {
 690                         CERROR("%s: fid init error: rc = %d\n",
 691                                m->opd_obd->obd_name, rc);
 692                         GOTO(out, rc);
 693                 }
 694         }
 695         /*
 696          * Initiate connect to OST
 697          */
 698         ll_generate_random_uuid(uuid);
 699         class_uuid_unparse(uuid, &m->opd_cluuid);
 700
 701         imp = obd->u.cli.cl_import;
 702
 703         rc = ptlrpc_init_import(imp);
 704         if (rc)
 705                 GOTO(out, rc);
 706         if (osdname)
 707                 OBD_FREE(osdname, MAX_OBD_NAME);
 708         RETURN(0);
 709
 710 out:
 711         if (!m->opd_connect_mdt)
 712                 /* stop sync thread */
 713                 osp_sync_fini(m);
 714 out_precreat:
 715         /* stop precreate thread */
 716         if (!m->opd_connect_mdt)
 717                 osp_precreate_fini(m);
 718 out_last_used:
 719         osp_last_used_fini(env, m);
 720 out_proc:
 721         ptlrpc_lprocfs_unregister_obd(obd);
 722         lprocfs_obd_cleanup(obd);
 723         class_destroy_import(obd->u.cli.cl_import);
 724         client_obd_cleanup(obd);
 725 out_ref:
 726         ptlrpcd_decref();
 727 out_disconnect:
 728         if (m->opd_connect_mdt) {
 729                 struct client_obd *cli = &m->opd_obd->u.cli;
 730                 if (cli->cl_rpc_lock != NULL) {
 731                         OBD_FREE_PTR(cli->cl_rpc_lock);
 732                         cli->cl_rpc_lock = NULL;
 733                 }
 734         }
 735         obd_disconnect(m->opd_storage_exp);
 736 out_fini:
 737         if (osdname)
 738                 OBD_FREE(osdname, MAX_OBD_NAME);
 739         RETURN(rc);
 740 }
 741
 742 static struct lu_device *osp_device_free(const struct lu_env *env,
 743                                          struct lu_device *lu)
 744 {
 745         struct osp_device *m = lu2osp_dev(lu);
 746
 747         ENTRY;
 748
 749         if (cfs_atomic_read(&lu->ld_ref) && lu->ld_site) {
 750                 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
 751                 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
 752         }
 753         dt_device_fini(&m->opd_dt_dev);
 754         OBD_FREE_PTR(m);
 755         RETURN(NULL);
 756 }
 757
 758 static struct lu_device *osp_device_alloc(const struct lu_env *env,
 759                                           struct lu_device_type *t,
 760                                           struct lustre_cfg *lcfg)
 761 {
 762         struct osp_device *m;
 763         struct lu_device  *l;
 764
 765         OBD_ALLOC_PTR(m);
 766         if (m == NULL) {
 767                 l = ERR_PTR(-ENOMEM);
 768         } else {
 769                 int rc;
 770
 771                 l = osp2lu_dev(m);
 772                 dt_device_init(&m->opd_dt_dev, t);
 773                 rc = osp_init0(env, m, t, lcfg);
 774                 if (rc != 0) {
 775                         osp_device_free(env, l);
 776                         l = ERR_PTR(rc);
 777                 }
 778         }
 779         return l;
 780 }
 781
 782 static struct lu_device *osp_device_fini(const struct lu_env *env,
 783                                          struct lu_device *d)
 784 {
 785         struct osp_device *m = lu2osp_dev(d);
 786         struct obd_import *imp;
 787         int                rc;
 788
 789         ENTRY;
 790
 791         if (m->opd_storage_exp)
 792                 obd_disconnect(m->opd_storage_exp);
 793
 794         imp = m->opd_obd->u.cli.cl_import;
 795
 796         if (imp->imp_rq_pool) {
 797                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 798                 imp->imp_rq_pool = NULL;
 799         }
 800
 801         obd_cleanup_client_import(m->opd_obd);
 802
 803         if (m->opd_symlink)
 804                 lprocfs_remove(&m->opd_symlink);
 805
 806         LASSERT(m->opd_obd);
 807         ptlrpc_lprocfs_unregister_obd(m->opd_obd);
 808         lprocfs_obd_cleanup(m->opd_obd);
 809
 810         if (m->opd_connect_mdt) {
 811                 struct client_obd *cli = &m->opd_obd->u.cli;
 812                 if (cli->cl_rpc_lock != NULL) {
 813                         OBD_FREE_PTR(cli->cl_rpc_lock);
 814                         cli->cl_rpc_lock = NULL;
 815                 }
 816         }
 817
 818         rc = client_obd_cleanup(m->opd_obd);
 819         LASSERTF(rc == 0, "error %d\n", rc);
 820
 821         ptlrpcd_decref();
 822
 823         RETURN(NULL);
 824 }
 825
 826 static int osp_reconnect(const struct lu_env *env,
 827                          struct obd_export *exp, struct obd_device *obd,
 828                          struct obd_uuid *cluuid,
 829                          struct obd_connect_data *data,
 830                          void *localdata)
 831 {
 832         return 0;
 833 }
 834
 835 /*
 836  * we use exports to track all LOD users
 837  */
 838 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
 839                            struct obd_device *obd, struct obd_uuid *cluuid,
 840                            struct obd_connect_data *data, void *localdata)
 841 {
 842         struct osp_device       *osp = lu2osp_dev(obd->obd_lu_dev);
 843         struct obd_connect_data *ocd;
 844         struct obd_import       *imp;
 845         struct lustre_handle     conn;
 846         int                      rc;
 847
 848         ENTRY;
 849
 850         CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
 851
 852         rc = class_connect(&conn, obd, cluuid);
 853         if (rc)
 854                 RETURN(rc);
 855
 856         *exp = class_conn2export(&conn);
 857         /* Why should there ever be more than 1 connect? */
 858         osp->opd_connects++;
 859         LASSERT(osp->opd_connects == 1);
 860
 861         osp->opd_exp = *exp;
 862
 863         imp = osp->opd_obd->u.cli.cl_import;
 864         imp->imp_dlm_handle = conn;
 865
 866         LASSERT(data != NULL);
 867         LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
 868         ocd = &imp->imp_connect_data;
 869         *ocd = *data;
 870
 871         imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
 872
 873         ocd->ocd_version = LUSTRE_VERSION_CODE;
 874         ocd->ocd_index = data->ocd_index;
 875         imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
 876
 877         rc = ptlrpc_connect_import(imp);
 878         if (rc) {
 879                 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
 880                 GOTO(out, rc);
 881         }
 882
 883         ptlrpc_pinger_add_import(imp);
 884
 885         if (osp->opd_connect_mdt && data->ocd_index == 0) {
 886                 /* set seq controller export for MDC0 if exists */
 887                 struct seq_server_site *ss;
 888
 889                 ss = lu_site2seq(osp2lu_dev(osp)->ld_site);
 890                 ss->ss_control_exp = class_export_get(*exp);
 891                 ss->ss_server_fld->lsf_control_exp = *exp;
 892         }
 893 out:
 894         RETURN(rc);
 895 }
 896
 897 /*
 898  * once last export (we don't count self-export) disappeared
 899  * osp can be released
 900  */
 901 static int osp_obd_disconnect(struct obd_export *exp)
 902 {
 903         struct obd_device *obd = exp->exp_obd;
 904         struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
 905         int                rc;
 906         ENTRY;
 907
 908         /* Only disconnect the underlying layers on the final disconnect. */
 909         LASSERT(osp->opd_connects == 1);
 910         osp->opd_connects--;
 911
 912         rc = class_disconnect(exp);
 913         if (rc) {
 914                 CERROR("%s: class disconnect error: rc = %d\n",
 915                        obd->obd_name, rc);
 916                 RETURN(rc);
 917         }
 918
 919         /* destroy the device */
 920         class_manual_cleanup(obd);
 921
 922         RETURN(rc);
 923 }
 924
 925 /*
 926  * lprocfs helpers still use OBD API, let's keep obd_statfs() support
 927  */
 928 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
 929                           struct obd_statfs *osfs, __u64 max_age, __u32 flags)
 930 {
 931         struct obd_statfs       *msfs;
 932         struct ptlrpc_request   *req;
 933         struct obd_import       *imp = NULL;
 934         int                      rc;
 935
 936         ENTRY;
 937
 938         /* Since the request might also come from lprocfs, so we need
 939          * sync this with client_disconnect_export Bug15684 */
 940         down_read(&exp->exp_obd->u.cli.cl_sem);
 941         if (exp->exp_obd->u.cli.cl_import)
 942                 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
 943         up_read(&exp->exp_obd->u.cli.cl_sem);
 944         if (!imp)
 945                 RETURN(-ENODEV);
 946
 947         /* We could possibly pass max_age in the request (as an absolute
 948          * timestamp or a "seconds.usec ago") so the target can avoid doing
 949          * extra calls into the filesystem if that isn't necessary (e.g.
 950          * during mount that would help a bit).  Having relative timestamps
 951          * is not so great if request processing is slow, while absolute
 952          * timestamps are not ideal because they need time synchronization. */
 953         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
 954
 955         class_import_put(imp);
 956
 957         if (req == NULL)
 958                 RETURN(-ENOMEM);
 959
 960         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
 961         if (rc) {
 962                 ptlrpc_request_free(req);
 963                 RETURN(rc);
 964         }
 965         ptlrpc_request_set_replen(req);
 966         req->rq_request_portal = OST_CREATE_PORTAL;
 967         ptlrpc_at_set_req_timeout(req);
 968
 969         if (flags & OBD_STATFS_NODELAY) {
 970                 /* procfs requests not want stat in wait for avoid deadlock */
 971                 req->rq_no_resend = 1;
 972                 req->rq_no_delay = 1;
 973         }
 974
 975         rc = ptlrpc_queue_wait(req);
 976         if (rc)
 977                 GOTO(out, rc);
 978
 979         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
 980         if (msfs == NULL)
 981                 GOTO(out, rc = -EPROTO);
 982
 983         *osfs = *msfs;
 984
 985         EXIT;
 986 out:
 987         ptlrpc_req_finished(req);
 988         return rc;
 989 }
 990
 991 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
 992                             enum obd_import_event event)
 993 {
 994         struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
 995
 996         switch (event) {
 997         case IMP_EVENT_DISCON:
 998                 d->opd_got_disconnected = 1;
 999                 d->opd_imp_connected = 0;
1000                 if (d->opd_connect_mdt)
1001                         break;
1002                 osp_pre_update_status(d, -ENODEV);
1003                 cfs_waitq_signal(&d->opd_pre_waitq);
1004                 CDEBUG(D_HA, "got disconnected\n");
1005                 break;
1006         case IMP_EVENT_INACTIVE:
1007                 d->opd_imp_active = 0;
1008                 if (d->opd_connect_mdt)
1009                         break;
1010                 osp_pre_update_status(d, -ENODEV);
1011                 cfs_waitq_signal(&d->opd_pre_waitq);
1012                 CDEBUG(D_HA, "got inactive\n");
1013                 break;
1014         case IMP_EVENT_ACTIVE:
1015                 d->opd_imp_active = 1;
1016                 if (d->opd_got_disconnected)
1017                         d->opd_new_connection = 1;
1018                 d->opd_imp_connected = 1;
1019                 d->opd_imp_seen_connected = 1;
1020                 if (d->opd_connect_mdt)
1021                         break;
1022                 cfs_waitq_signal(&d->opd_pre_waitq);
1023                 __osp_sync_check_for_work(d);
1024                 CDEBUG(D_HA, "got connected\n");
1025                 break;
1026         case IMP_EVENT_INVALIDATE:
1027                 if (obd->obd_namespace == NULL)
1028                         break;
1029                 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1030                 break;
1031         case IMP_EVENT_OCD:
1032         case IMP_EVENT_DEACTIVATE:
1033         case IMP_EVENT_ACTIVATE:
1034                 break;
1035         default:
1036                 CERROR("%s: unsupported import event: %#x\n",
1037                        obd->obd_name, event);
1038         }
1039         return 0;
1040 }
1041
1042 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1043                          void *karg, void *uarg)
1044 {
1045         struct obd_device       *obd = exp->exp_obd;
1046         struct osp_device       *d;
1047         struct obd_ioctl_data   *data = karg;
1048         int                      rc = 0;
1049
1050         ENTRY;
1051
1052         LASSERT(obd->obd_lu_dev);
1053         d = lu2osp_dev(obd->obd_lu_dev);
1054         LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1055
1056         if (!cfs_try_module_get(THIS_MODULE)) {
1057                 CERROR("%s: can't get module. Is it alive?", obd->obd_name);
1058                 return -EINVAL;
1059         }
1060
1061         switch (cmd) {
1062         case OBD_IOC_CLIENT_RECOVER:
1063                 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1064                                            data->ioc_inlbuf1, 0);
1065                 if (rc > 0)
1066                         rc = 0;
1067                 break;
1068         case IOC_OSC_SET_ACTIVE:
1069                 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1070                                               data->ioc_offset);
1071                 break;
1072         case OBD_IOC_PING_TARGET:
1073                 rc = ptlrpc_obd_ping(obd);
1074                 break;
1075         default:
1076                 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1077                        cmd, cfs_curproc_comm());
1078                 rc = -ENOTTY;
1079         }
1080         cfs_module_put(THIS_MODULE);
1081         return rc;
1082 }
1083
1084 static int osp_obd_health_check(const struct lu_env *env,
1085                                 struct obd_device *obd)
1086 {
1087         struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1088
1089         ENTRY;
1090
1091         /*
1092          * 1.8/2.0 behaviour is that OST being connected once at least
1093          * is considired "healthy". and one "healty" OST is enough to
1094          * allow lustre clients to connect to MDS
1095          */
1096         LASSERT(d);
1097         RETURN(!d->opd_imp_seen_connected);
1098 }
1099
1100 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1101 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1102 static void osp_key_exit(const struct lu_context *ctx,
1103                          struct lu_context_key *key, void *data)
1104 {
1105         struct osp_thread_info *info = data;
1106
1107         info->osi_attr.la_valid = 0;
1108 }
1109
1110 struct lu_context_key osp_thread_key = {
1111         .lct_tags = LCT_MD_THREAD,
1112         .lct_init = osp_key_init,
1113         .lct_fini = osp_key_fini,
1114         .lct_exit = osp_key_exit
1115 };
1116
1117 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1118 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1119
1120 struct lu_context_key osp_txn_key = {
1121         .lct_tags = LCT_OSP_THREAD,
1122         .lct_init = osp_txn_key_init,
1123         .lct_fini = osp_txn_key_fini
1124 };
1125 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1126
1127 static struct lu_device_type_operations osp_device_type_ops = {
1128         .ldto_init           = osp_type_init,
1129         .ldto_fini           = osp_type_fini,
1130
1131         .ldto_start          = osp_type_start,
1132         .ldto_stop           = osp_type_stop,
1133
1134         .ldto_device_alloc   = osp_device_alloc,
1135         .ldto_device_free    = osp_device_free,
1136
1137         .ldto_device_fini    = osp_device_fini
1138 };
1139
1140 static struct lu_device_type osp_device_type = {
1141         .ldt_tags     = LU_DEVICE_DT,
1142         .ldt_name     = LUSTRE_OSP_NAME,
1143         .ldt_ops      = &osp_device_type_ops,
1144         .ldt_ctx_tags = LCT_MD_THREAD
1145 };
1146
1147 static struct obd_ops osp_obd_device_ops = {
1148         .o_owner        = THIS_MODULE,
1149         .o_add_conn     = client_import_add_conn,
1150         .o_del_conn     = client_import_del_conn,
1151         .o_reconnect    = osp_reconnect,
1152         .o_connect      = osp_obd_connect,
1153         .o_disconnect   = osp_obd_disconnect,
1154         .o_health_check = osp_obd_health_check,
1155         .o_import_event = osp_import_event,
1156         .o_iocontrol    = osp_iocontrol,
1157         .o_statfs       = osp_obd_statfs,
1158         .o_fid_init     = client_fid_init,
1159         .o_fid_fini     = client_fid_fini,
1160 };
1161
1162 struct llog_operations osp_mds_ost_orig_logops;
1163
1164 static int __init osp_mod_init(void)
1165 {
1166         struct lprocfs_static_vars       lvars;
1167         cfs_proc_dir_entry_t            *osc_proc_dir;
1168         int                              rc;
1169
1170         rc = lu_kmem_init(osp_caches);
1171         if (rc)
1172                 return rc;
1173
1174         lprocfs_osp_init_vars(&lvars);
1175
1176         rc = class_register_type(&osp_obd_device_ops, NULL, lvars.module_vars,
1177                                  LUSTRE_OSP_NAME, &osp_device_type);
1178
1179         /* create "osc" entry in procfs for compatibility purposes */
1180         if (rc != 0) {
1181                 lu_kmem_fini(osp_caches);
1182                 return rc;
1183         }
1184
1185         lprocfs_lwp_init_vars(&lvars);
1186
1187         rc = class_register_type(&lwp_obd_device_ops, NULL, lvars.module_vars,
1188                                  LUSTRE_LWP_NAME, &lwp_device_type);
1189         if (rc != 0) {
1190                 class_unregister_type(LUSTRE_OSP_NAME);
1191                 lu_kmem_fini(osp_caches);
1192                 return rc;
1193         }
1194
1195         /* Note: add_rec/delcare_add_rec will be only used by catalogs */
1196         osp_mds_ost_orig_logops = llog_osd_ops;
1197         osp_mds_ost_orig_logops.lop_add = llog_cat_add_rec;
1198         osp_mds_ost_orig_logops.lop_declare_add = llog_cat_declare_add_rec;
1199
1200         osc_proc_dir = lprocfs_srch(proc_lustre_root, "osc");
1201         if (osc_proc_dir == NULL) {
1202                 osc_proc_dir = lprocfs_register("osc", proc_lustre_root, NULL,
1203                                                 NULL);
1204                 if (IS_ERR(osc_proc_dir))
1205                         CERROR("osp: can't create compat entry \"osc\": %d\n",
1206                                (int) PTR_ERR(osc_proc_dir));
1207         }
1208         return rc;
1209 }
1210
1211 static void __exit osp_mod_exit(void)
1212 {
1213         lprocfs_try_remove_proc_entry("osc", proc_lustre_root);
1214
1215         class_unregister_type(LUSTRE_LWP_NAME);
1216         class_unregister_type(LUSTRE_OSP_NAME);
1217         lu_kmem_fini(osp_caches);
1218 }
1219
1220 MODULE_AUTHOR("Intel, Inc. <http://www.intel.com/>");
1221 MODULE_DESCRIPTION("Lustre OST Proxy Device ("LUSTRE_OSP_NAME")");
1222 MODULE_LICENSE("GPL");
1223
1224 cfs_module(osp, LUSTRE_VERSION_STRING, osp_mod_init, osp_mod_exit);