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) 2011, 2012, Intel, Inc.
  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  */
  43
  44 #ifndef EXPORT_SYMTAB
  45 # define EXPORT_SYMTAB
  46 #endif
  47 #define DEBUG_SUBSYSTEM S_MDS
  48
  49 #include <obd_class.h>
  50 #include <lustre_param.h>
  51
  52 #include "osp_internal.h"
  53
  54 /* Slab for OSP object allocation */
  55 cfs_mem_cache_t *osp_object_kmem;
  56
  57 static struct lu_kmem_descr osp_caches[] = {
  58         {
  59                 .ckd_cache = &osp_object_kmem,
  60                 .ckd_name  = "osp_obj",
  61                 .ckd_size  = sizeof(struct osp_object)
  62         },
  63         {
  64                 .ckd_cache = NULL
  65         }
  66 };
  67
  68 struct lu_object *osp_object_alloc(const struct lu_env *env,
  69                                    const struct lu_object_header *hdr,
  70                                    struct lu_device *d)
  71 {
  72         struct lu_object_header *h;
  73         struct osp_object       *o;
  74         struct lu_object        *l;
  75
  76         LASSERT(hdr == NULL);
  77
  78         OBD_SLAB_ALLOC_PTR_GFP(o, osp_object_kmem, CFS_ALLOC_IO);
  79         if (o != NULL) {
  80                 l = &o->opo_obj.do_lu;
  81                 h = &o->opo_header;
  82
  83                 lu_object_header_init(h);
  84                 dt_object_init(&o->opo_obj, h, d);
  85                 lu_object_add_top(h, l);
  86
  87                 l->lo_ops = &osp_lu_obj_ops;
  88
  89                 return l;
  90         } else {
  91                 return NULL;
  92         }
  93 }
  94
  95 /* Update opd_last_used_id along with checking for gap in objid sequence */
  96 void osp_update_last_id(struct osp_device *d, obd_id objid)
  97 {
  98         /*
  99          * we might have lost precreated objects due to VBR and precreate
 100          * orphans, the gap in objid can be calculated properly only here
 101          */
 102         if (objid > le64_to_cpu(d->opd_last_used_id)) {
 103                 if (objid - le64_to_cpu(d->opd_last_used_id) > 1) {
 104                         d->opd_gap_start = le64_to_cpu(d->opd_last_used_id) + 1;
 105                         d->opd_gap_count = objid - d->opd_gap_start;
 106                         CDEBUG(D_HA, "Gap in objids: %d, start = %llu\n",
 107                                d->opd_gap_count, d->opd_gap_start);
 108                 }
 109                 d->opd_last_used_id = cpu_to_le64(objid);
 110         }
 111 }
 112
 113 static int osp_last_used_init(const struct lu_env *env, struct osp_device *m)
 114 {
 115         struct osp_thread_info  *osi = osp_env_info(env);
 116         struct dt_object_format  dof = { 0 };
 117         struct dt_object        *o;
 118         int                      rc;
 119
 120         ENTRY;
 121
 122         osi->osi_attr.la_valid = LA_MODE;
 123         osi->osi_attr.la_mode = S_IFREG | 0644;
 124         lu_local_obj_fid(&osi->osi_fid, MDD_LOV_OBJ_OID);
 125         dof.dof_type = DFT_REGULAR;
 126         o = dt_find_or_create(env, m->opd_storage, &osi->osi_fid, &dof,
 127                               &osi->osi_attr);
 128         if (IS_ERR(o))
 129                 RETURN(PTR_ERR(o));
 130
 131         rc = dt_attr_get(env, o, &osi->osi_attr, NULL);
 132         if (rc)
 133                 GOTO(out, rc);
 134
 135         /* object will be released in device cleanup path */
 136         m->opd_last_used_file = o;
 137
 138         if (osi->osi_attr.la_size >= sizeof(osi->osi_id) *
 139                                      (m->opd_index + 1)) {
 140                 osp_objid_buf_prep(osi, m, m->opd_index);
 141                 rc = dt_record_read(env, o, &osi->osi_lb, &osi->osi_off);
 142                 if (rc != 0)
 143                         GOTO(out, rc);
 144         } else {
 145                 /* reset value to 0, just to make sure and change file's size */
 146                 struct thandle *th;
 147
 148                 m->opd_last_used_id = 0;
 149                 osp_objid_buf_prep(osi, m, m->opd_index);
 150
 151                 th = dt_trans_create(env, m->opd_storage);
 152                 if (IS_ERR(th))
 153                         GOTO(out, rc = PTR_ERR(th));
 154
 155                 rc = dt_declare_record_write(env, m->opd_last_used_file,
 156                                              osi->osi_lb.lb_len, osi->osi_off,
 157                                              th);
 158                 if (rc) {
 159                         dt_trans_stop(env, m->opd_storage, th);
 160                         GOTO(out, rc);
 161                 }
 162
 163                 rc = dt_trans_start_local(env, m->opd_storage, th);
 164                 if (rc) {
 165                         dt_trans_stop(env, m->opd_storage, th);
 166                         GOTO(out, rc);
 167                 }
 168
 169                 rc = dt_record_write(env, m->opd_last_used_file, &osi->osi_lb,
 170                                      &osi->osi_off, th);
 171                 dt_trans_stop(env, m->opd_storage, th);
 172                 if (rc)
 173                         GOTO(out, rc);
 174         }
 175         CDEBUG(D_HA, "%s: Read last used ID: "LPU64"\n", m->opd_obd->obd_name,
 176                le64_to_cpu(m->opd_last_used_id));
 177         RETURN(0);
 178 out:
 179         CERROR("%s: can't initialize lov_objid: %d\n",
 180                m->opd_obd->obd_name, rc);
 181         lu_object_put(env, &o->do_lu);
 182         m->opd_last_used_file = NULL;
 183         return rc;
 184 }
 185
 186 static void osp_last_used_fini(const struct lu_env *env, struct osp_device *d)
 187 {
 188         if (d->opd_last_used_file != NULL) {
 189                 lu_object_put(env, &d->opd_last_used_file->do_lu);
 190                 d->opd_last_used_file = NULL;
 191         }
 192 }
 193
 194 int osp_disconnect(struct osp_device *d)
 195 {
 196         struct obd_import *imp;
 197         int rc = 0;
 198
 199         imp = d->opd_obd->u.cli.cl_import;
 200
 201         /* Mark import deactivated now, so we don't try to reconnect if any
 202          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 203          * fully deactivate the import, or that would drop all requests. */
 204         LASSERT(imp != NULL);
 205         spin_lock(&imp->imp_lock);
 206         imp->imp_deactive = 1;
 207         spin_unlock(&imp->imp_lock);
 208
 209         ptlrpc_deactivate_import(imp);
 210
 211         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 212          * delete it regardless.  (It's safe to delete an import that was
 213          * never added.) */
 214         (void)ptlrpc_pinger_del_import(imp);
 215
 216         rc = ptlrpc_disconnect_import(imp, 0);
 217         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
 218                 rc = 0;
 219         if (rc)
 220                 CERROR("%s: can't disconnect: rc = %d\n",
 221                        d->opd_obd->obd_name, rc);
 222
 223         ptlrpc_invalidate_import(imp);
 224
 225         RETURN(rc);
 226 }
 227
 228 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
 229 {
 230         int                      rc = 0;
 231         ENTRY;
 232
 233         if (is_osp_on_ost(d->opd_obd->obd_name)) {
 234                 rc = osp_disconnect(d);
 235                 RETURN(rc);
 236         }
 237
 238         LASSERT(env);
 239         /* release last_used file */
 240         osp_last_used_fini(env, d);
 241
 242         rc = osp_disconnect(d);
 243
 244         /* stop precreate thread */
 245         osp_precreate_fini(d);
 246
 247         /* stop sync thread */
 248         osp_sync_fini(d);
 249
 250         RETURN(rc);
 251 }
 252
 253 static int osp_process_config(const struct lu_env *env,
 254                               struct lu_device *dev, struct lustre_cfg *lcfg)
 255 {
 256         struct osp_device               *d = lu2osp_dev(dev);
 257         struct lprocfs_static_vars       lvars = { 0 };
 258         int                              rc;
 259
 260         ENTRY;
 261
 262         switch (lcfg->lcfg_command) {
 263         case LCFG_CLEANUP:
 264                 if (!is_osp_on_ost(d->opd_obd->obd_name))
 265                         lu_dev_del_linkage(dev->ld_site, dev);
 266                 rc = osp_shutdown(env, d);
 267                 break;
 268         case LCFG_PARAM:
 269                 lprocfs_osp_init_vars(&lvars);
 270
 271                 LASSERT(d->opd_obd);
 272                 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
 273                                               lcfg, d->opd_obd);
 274                 if (rc > 0)
 275                         rc = 0;
 276                 if (rc == -ENOSYS) {
 277                         /* class_process_proc_param() haven't found matching
 278                          * parameter and returned ENOSYS so that layer(s)
 279                          * below could use that. But OSP is the bottom, so
 280                          * just ignore it */
 281                         CERROR("%s: unknown param %s\n",
 282                                (char *)lustre_cfg_string(lcfg, 0),
 283                                (char *)lustre_cfg_string(lcfg, 1));
 284                         rc = 0;
 285                 }
 286                 break;
 287         default:
 288                 CERROR("%s: unknown command %u\n",
 289                        (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
 290                 rc = 0;
 291                 break;
 292         }
 293
 294         RETURN(rc);
 295 }
 296
 297 static int osp_recovery_complete(const struct lu_env *env,
 298                                  struct lu_device *dev)
 299 {
 300         struct osp_device       *osp = lu2osp_dev(dev);
 301         int                      rc = 0;
 302
 303         ENTRY;
 304         osp->opd_recovery_completed = 1;
 305         cfs_waitq_signal(&osp->opd_pre_waitq);
 306         RETURN(rc);
 307 }
 308
 309 const struct lu_device_operations osp_lu_ops = {
 310         .ldo_object_alloc       = osp_object_alloc,
 311         .ldo_process_config     = osp_process_config,
 312         .ldo_recovery_complete  = osp_recovery_complete,
 313 };
 314
 315 /**
 316  * provides with statfs from corresponded OST
 317  *
 318  */
 319 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
 320                       struct obd_statfs *sfs)
 321 {
 322         struct osp_device *d = dt2osp_dev(dev);
 323
 324         ENTRY;
 325
 326         if (unlikely(d->opd_imp_active == 0))
 327                 RETURN(-ENOTCONN);
 328
 329         /* return recently updated data */
 330         *sfs = d->opd_statfs;
 331
 332         /*
 333          * layer above osp (usually lod) can use ffree to estimate
 334          * how many objects are available for immediate creation
 335          */
 336         spin_lock(&d->opd_pre_lock);
 337         sfs->os_fprecreated = d->opd_pre_last_created - d->opd_pre_used_id;
 338         sfs->os_fprecreated -= d->opd_pre_reserved;
 339         spin_unlock(&d->opd_pre_lock);
 340
 341         LASSERT(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2);
 342
 343         CDEBUG(D_OTHER, "%s: "LPU64" blocks, "LPU64" free, "LPU64" avail, "
 344                LPU64" files, "LPU64" free files\n", d->opd_obd->obd_name,
 345                sfs->os_blocks, sfs->os_bfree, sfs->os_bavail,
 346                sfs->os_files, sfs->os_ffree);
 347         RETURN(0);
 348 }
 349
 350 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
 351 {
 352         ENTRY;
 353
 354         /*
 355          * XXX: wake up sync thread, command it to start flushing asap?
 356          */
 357
 358         RETURN(0);
 359 }
 360
 361 const struct dt_device_operations osp_dt_ops = {
 362         .dt_statfs      = osp_statfs,
 363         .dt_sync        = osp_sync,
 364 };
 365
 366 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *m,
 367                               const char *nextdev)
 368 {
 369         struct obd_connect_data *data = NULL;
 370         struct obd_device       *obd;
 371         int                      rc;
 372
 373         ENTRY;
 374
 375         LASSERT(m->opd_storage_exp == NULL);
 376
 377         OBD_ALLOC_PTR(data);
 378         if (data == NULL)
 379                 RETURN(-ENOMEM);
 380
 381         obd = class_name2obd(nextdev);
 382         if (obd == NULL) {
 383                 CERROR("%s: can't locate next device: %s\n",
 384                        m->opd_obd->obd_name, nextdev);
 385                 GOTO(out, rc = -ENOTCONN);
 386         }
 387
 388         rc = obd_connect(env, &m->opd_storage_exp, obd, &obd->obd_uuid, data,
 389                          NULL);
 390         if (rc) {
 391                 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
 392                        m->opd_obd->obd_name, nextdev, rc);
 393                 GOTO(out, rc);
 394         }
 395
 396         m->opd_dt_dev.dd_lu_dev.ld_site =
 397                 m->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
 398         LASSERT(m->opd_dt_dev.dd_lu_dev.ld_site);
 399         m->opd_storage = lu2dt_dev(m->opd_storage_exp->exp_obd->obd_lu_dev);
 400
 401 out:
 402         OBD_FREE_PTR(data);
 403         RETURN(rc);
 404 }
 405
 406 static int osp_init0(const struct lu_env *env, struct osp_device *m,
 407                      struct lu_device_type *ldt, struct lustre_cfg *cfg)
 408 {
 409         struct obd_device               *obd;
 410         struct obd_import               *imp;
 411         class_uuid_t                     uuid;
 412         char                            *src, *ost, *mdt, *osdname = NULL;
 413         int                              rc, idx;
 414
 415         ENTRY;
 416
 417         obd = class_name2obd(lustre_cfg_string(cfg, 0));
 418         if (obd == NULL) {
 419                 CERROR("Cannot find obd with name %s\n",
 420                        lustre_cfg_string(cfg, 0));
 421                 RETURN(-ENODEV);
 422         }
 423         m->opd_obd = obd;
 424
 425         /* There is no record in the MDT configuration for the local disk
 426          * device, so we have to extract this from elsewhere in the profile.
 427          * The only information we get at setup is from the OSC records:
 428          * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 429          * Note that 1.8 generated configs are missing the -MDTxxxx part.
 430          * We need to reconstruct the name of the underlying OSD from this:
 431          * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".  We
 432          * also need to determine the OST index from this - will be used
 433          * to calculate the offset in shared lov_objids file later */
 434
 435         src = lustre_cfg_string(cfg, 0);
 436         if (src == NULL)
 437                 RETURN(-EINVAL);
 438
 439         ost = strstr(src, "-OST");
 440         if (ost == NULL)
 441                 RETURN(-EINVAL);
 442
 443         idx = simple_strtol(ost + 4, &mdt, 16);
 444         if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
 445                 CERROR("%s: invalid OST index in '%s'\n", obd->obd_name, src);
 446                 GOTO(out_fini, rc = -EINVAL);
 447         }
 448         m->opd_index = idx;
 449
 450         idx = ost - src;
 451         /* check the fsname length, and after this everything else will fit */
 452         if (idx > MTI_NAME_MAXLEN) {
 453                 CERROR("%s: fsname too long in '%s'\n", obd->obd_name, src);
 454                 GOTO(out_fini, rc = -EINVAL);
 455         }
 456
 457         OBD_ALLOC(osdname, MAX_OBD_NAME);
 458         if (osdname == NULL)
 459                 GOTO(out_fini, rc = -ENOMEM);
 460
 461         memcpy(osdname, src, idx); /* copy just the fsname part */
 462         osdname[idx] = '\0';
 463
 464         mdt = strstr(mdt, "-MDT");
 465         if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
 466                 strcat(osdname, "-MDT0000");
 467         else
 468                 strcat(osdname, mdt);
 469         strcat(osdname, "-osd");
 470         CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
 471
 472         m->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
 473         m->opd_dt_dev.dd_ops = &osp_dt_ops;
 474         obd->obd_lu_dev = &m->opd_dt_dev.dd_lu_dev;
 475
 476         rc = osp_connect_to_osd(env, m, osdname);
 477         if (rc)
 478                 GOTO(out_fini, rc);
 479
 480         rc = ptlrpcd_addref();
 481         if (rc)
 482                 GOTO(out_disconnect, rc);
 483
 484         rc = client_obd_setup(obd, cfg);
 485         if (rc) {
 486                 CERROR("%s: can't setup obd: %d\n", m->opd_obd->obd_name, rc);
 487                 GOTO(out_ref, rc);
 488         }
 489
 490         osp_lprocfs_init(m);
 491
 492         /*
 493          * Initialize last id from the storage - will be used in orphan cleanup
 494          */
 495         rc = osp_last_used_init(env, m);
 496         if (rc)
 497                 GOTO(out_proc, rc);
 498
 499         /*
 500          * Initialize precreation thread, it handles new connections as well
 501          */
 502         rc = osp_init_precreate(m);
 503         if (rc)
 504                 GOTO(out_last_used, rc);
 505
 506         /*
 507          * Initialize synhronization mechanism taking care of propogating
 508          * changes to OST in near transactional manner
 509          */
 510         rc = osp_sync_init(env, m);
 511         if (rc)
 512                 GOTO(out_precreat, rc);
 513
 514         /*
 515          * Initiate connect to OST
 516          */
 517         ll_generate_random_uuid(uuid);
 518         class_uuid_unparse(uuid, &m->opd_cluuid);
 519
 520         imp = obd->u.cli.cl_import;
 521
 522         rc = ptlrpc_init_import(imp);
 523         if (rc)
 524                 GOTO(out, rc);
 525         if (osdname)
 526                 OBD_FREE(osdname, MAX_OBD_NAME);
 527         RETURN(0);
 528
 529 out:
 530         /* stop sync thread */
 531         osp_sync_fini(m);
 532 out_precreat:
 533         /* stop precreate thread */
 534         osp_precreate_fini(m);
 535 out_last_used:
 536         osp_last_used_fini(env, m);
 537 out_proc:
 538         ptlrpc_lprocfs_unregister_obd(obd);
 539         lprocfs_obd_cleanup(obd);
 540         class_destroy_import(obd->u.cli.cl_import);
 541         client_obd_cleanup(obd);
 542 out_ref:
 543         ptlrpcd_decref();
 544 out_disconnect:
 545         obd_disconnect(m->opd_storage_exp);
 546 out_fini:
 547         if (osdname)
 548                 OBD_FREE(osdname, MAX_OBD_NAME);
 549         RETURN(rc);
 550 }
 551
 552 static struct lu_device *osp_device_free(const struct lu_env *env,
 553                                          struct lu_device *lu)
 554 {
 555         struct osp_device *m = lu2osp_dev(lu);
 556
 557         ENTRY;
 558
 559         if (cfs_atomic_read(&lu->ld_ref) && lu->ld_site) {
 560                 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
 561                 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
 562         }
 563         dt_device_fini(&m->opd_dt_dev);
 564         OBD_FREE_PTR(m);
 565         RETURN(NULL);
 566 }
 567
 568 static struct lu_device *osp_device_alloc(const struct lu_env *env,
 569                                           struct lu_device_type *t,
 570                                           struct lustre_cfg *lcfg)
 571 {
 572         struct osp_device *m;
 573         struct lu_device  *l;
 574
 575         OBD_ALLOC_PTR(m);
 576         if (m == NULL) {
 577                 l = ERR_PTR(-ENOMEM);
 578         } else {
 579                 int rc;
 580
 581                 l = osp2lu_dev(m);
 582                 dt_device_init(&m->opd_dt_dev, t);
 583                 if (is_osp_on_ost(lustre_cfg_string(lcfg, 0)))
 584                         rc = osp_init_for_ost(env, m, t, lcfg);
 585                 else
 586                         rc = osp_init0(env, m, t, lcfg);
 587                 if (rc != 0) {
 588                         osp_device_free(env, l);
 589                         l = ERR_PTR(rc);
 590                 }
 591         }
 592         return l;
 593 }
 594
 595 static struct lu_device *osp_device_fini(const struct lu_env *env,
 596                                          struct lu_device *d)
 597 {
 598         struct osp_device *m = lu2osp_dev(d);
 599         struct obd_import *imp;
 600         int                rc;
 601
 602         ENTRY;
 603
 604         if (m->opd_storage_exp)
 605                 obd_disconnect(m->opd_storage_exp);
 606
 607         if (is_osp_on_ost(m->opd_obd->obd_name))
 608                 osp_fini_for_ost(m);
 609
 610         imp = m->opd_obd->u.cli.cl_import;
 611
 612         if (imp->imp_rq_pool) {
 613                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 614                 imp->imp_rq_pool = NULL;
 615         }
 616
 617         obd_cleanup_client_import(m->opd_obd);
 618
 619         if (m->opd_symlink)
 620                 lprocfs_remove(&m->opd_symlink);
 621
 622         LASSERT(m->opd_obd);
 623         ptlrpc_lprocfs_unregister_obd(m->opd_obd);
 624         lprocfs_obd_cleanup(m->opd_obd);
 625
 626         rc = client_obd_cleanup(m->opd_obd);
 627         LASSERTF(rc == 0, "error %d\n", rc);
 628
 629         ptlrpcd_decref();
 630
 631         RETURN(NULL);
 632 }
 633
 634 static int osp_reconnect(const struct lu_env *env,
 635                          struct obd_export *exp, struct obd_device *obd,
 636                          struct obd_uuid *cluuid,
 637                          struct obd_connect_data *data,
 638                          void *localdata)
 639 {
 640         return 0;
 641 }
 642
 643 /*
 644  * we use exports to track all LOD users
 645  */
 646 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
 647                            struct obd_device *obd, struct obd_uuid *cluuid,
 648                            struct obd_connect_data *data, void *localdata)
 649 {
 650         struct osp_device       *osp = lu2osp_dev(obd->obd_lu_dev);
 651         struct obd_connect_data *ocd;
 652         struct obd_import       *imp;
 653         struct lustre_handle     conn;
 654         int                      rc;
 655
 656         ENTRY;
 657
 658         CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
 659
 660         rc = class_connect(&conn, obd, cluuid);
 661         if (rc)
 662                 RETURN(rc);
 663
 664         *exp = class_conn2export(&conn);
 665         if (is_osp_on_ost(obd->obd_name))
 666                 osp->opd_exp = *exp;
 667
 668         /* Why should there ever be more than 1 connect? */
 669         osp->opd_connects++;
 670         LASSERT(osp->opd_connects == 1);
 671
 672         imp = osp->opd_obd->u.cli.cl_import;
 673         imp->imp_dlm_handle = conn;
 674
 675         ocd = &imp->imp_connect_data;
 676         ocd->ocd_connect_flags = OBD_CONNECT_AT |
 677                                  OBD_CONNECT_FULL20 |
 678                                  OBD_CONNECT_INDEX |
 679 #ifdef HAVE_LRU_RESIZE_SUPPORT
 680                                  OBD_CONNECT_LRU_RESIZE |
 681 #endif
 682                                  OBD_CONNECT_MDS |
 683                                  OBD_CONNECT_OSS_CAPA |
 684                                  OBD_CONNECT_REQPORTAL |
 685                                  OBD_CONNECT_SKIP_ORPHAN |
 686                                  OBD_CONNECT_VERSION |
 687                                  OBD_CONNECT_FID |
 688                                  OBD_CONNECT_LVB_TYPE;
 689
 690         if (is_osp_on_ost(osp->opd_obd->obd_name))
 691                 ocd->ocd_connect_flags |= OBD_CONNECT_LIGHTWEIGHT;
 692
 693         ocd->ocd_version = LUSTRE_VERSION_CODE;
 694         LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
 695         ocd->ocd_index = data->ocd_index;
 696         imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
 697
 698         rc = ptlrpc_connect_import(imp);
 699         if (rc) {
 700                 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
 701                 GOTO(out, rc);
 702         }
 703
 704         ptlrpc_pinger_add_import(imp);
 705
 706 out:
 707         RETURN(rc);
 708 }
 709
 710 /*
 711  * once last export (we don't count self-export) disappeared
 712  * osp can be released
 713  */
 714 static int osp_obd_disconnect(struct obd_export *exp)
 715 {
 716         struct obd_device *obd = exp->exp_obd;
 717         struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
 718         int                rc;
 719         ENTRY;
 720
 721         /* Only disconnect the underlying layers on the final disconnect. */
 722         LASSERT(osp->opd_connects == 1);
 723         osp->opd_connects--;
 724
 725         rc = class_disconnect(exp);
 726         if (rc) {
 727                 CERROR("%s: class disconnect error: rc = %d\n",
 728                        obd->obd_name, rc);
 729                 RETURN(rc);
 730         }
 731
 732         /* destroy the device */
 733         if (!is_osp_on_ost(obd->obd_name))
 734                 class_manual_cleanup(obd);
 735
 736         RETURN(rc);
 737 }
 738
 739 /*
 740  * lprocfs helpers still use OBD API, let's keep obd_statfs() support
 741  */
 742 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
 743                           struct obd_statfs *osfs, __u64 max_age, __u32 flags)
 744 {
 745         struct obd_statfs       *msfs;
 746         struct ptlrpc_request   *req;
 747         struct obd_import       *imp = NULL;
 748         int                      rc;
 749
 750         ENTRY;
 751
 752         /* Since the request might also come from lprocfs, so we need
 753          * sync this with client_disconnect_export Bug15684 */
 754         down_read(&exp->exp_obd->u.cli.cl_sem);
 755         if (exp->exp_obd->u.cli.cl_import)
 756                 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
 757         up_read(&exp->exp_obd->u.cli.cl_sem);
 758         if (!imp)
 759                 RETURN(-ENODEV);
 760
 761         /* We could possibly pass max_age in the request (as an absolute
 762          * timestamp or a "seconds.usec ago") so the target can avoid doing
 763          * extra calls into the filesystem if that isn't necessary (e.g.
 764          * during mount that would help a bit).  Having relative timestamps
 765          * is not so great if request processing is slow, while absolute
 766          * timestamps are not ideal because they need time synchronization. */
 767         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
 768
 769         class_import_put(imp);
 770
 771         if (req == NULL)
 772                 RETURN(-ENOMEM);
 773
 774         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
 775         if (rc) {
 776                 ptlrpc_request_free(req);
 777                 RETURN(rc);
 778         }
 779         ptlrpc_request_set_replen(req);
 780         req->rq_request_portal = OST_CREATE_PORTAL;
 781         ptlrpc_at_set_req_timeout(req);
 782
 783         if (flags & OBD_STATFS_NODELAY) {
 784                 /* procfs requests not want stat in wait for avoid deadlock */
 785                 req->rq_no_resend = 1;
 786                 req->rq_no_delay = 1;
 787         }
 788
 789         rc = ptlrpc_queue_wait(req);
 790         if (rc)
 791                 GOTO(out, rc);
 792
 793         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
 794         if (msfs == NULL)
 795                 GOTO(out, rc = -EPROTO);
 796
 797         *osfs = *msfs;
 798
 799         EXIT;
 800 out:
 801         ptlrpc_req_finished(req);
 802         return rc;
 803 }
 804
 805 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
 806                             enum obd_import_event event)
 807 {
 808         struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
 809
 810         switch (event) {
 811         case IMP_EVENT_DISCON:
 812                 d->opd_got_disconnected = 1;
 813                 d->opd_imp_connected = 0;
 814                 if (is_osp_on_ost(d->opd_obd->obd_name))
 815                         break;
 816                 osp_pre_update_status(d, -ENODEV);
 817                 cfs_waitq_signal(&d->opd_pre_waitq);
 818                 CDEBUG(D_HA, "got disconnected\n");
 819                 break;
 820         case IMP_EVENT_INACTIVE:
 821                 d->opd_imp_active = 0;
 822                 if (is_osp_on_ost(d->opd_obd->obd_name))
 823                         break;
 824                 osp_pre_update_status(d, -ENODEV);
 825                 cfs_waitq_signal(&d->opd_pre_waitq);
 826                 CDEBUG(D_HA, "got inactive\n");
 827                 break;
 828         case IMP_EVENT_ACTIVE:
 829                 d->opd_imp_active = 1;
 830                 if (d->opd_got_disconnected)
 831                         d->opd_new_connection = 1;
 832                 d->opd_imp_connected = 1;
 833                 d->opd_imp_seen_connected = 1;
 834                 if (is_osp_on_ost(d->opd_obd->obd_name))
 835                         break;
 836                 cfs_waitq_signal(&d->opd_pre_waitq);
 837                 __osp_sync_check_for_work(d);
 838                 CDEBUG(D_HA, "got connected\n");
 839                 break;
 840         case IMP_EVENT_INVALIDATE:
 841                 if (obd->obd_namespace == NULL)
 842                         break;
 843                 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
 844                 break;
 845         case IMP_EVENT_OCD:
 846         case IMP_EVENT_DEACTIVATE:
 847         case IMP_EVENT_ACTIVATE:
 848                 break;
 849         default:
 850                 CERROR("%s: unsupported import event: %#x\n",
 851                        obd->obd_name, event);
 852         }
 853         return 0;
 854 }
 855
 856 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
 857                          void *karg, void *uarg)
 858 {
 859         struct obd_device       *obd = exp->exp_obd;
 860         struct osp_device       *d;
 861         struct obd_ioctl_data   *data = karg;
 862         int                      rc = 0;
 863
 864         ENTRY;
 865
 866         LASSERT(obd->obd_lu_dev);
 867         d = lu2osp_dev(obd->obd_lu_dev);
 868         LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
 869
 870         if (!cfs_try_module_get(THIS_MODULE)) {
 871                 CERROR("%s: can't get module. Is it alive?", obd->obd_name);
 872                 return -EINVAL;
 873         }
 874
 875         switch (cmd) {
 876         case OBD_IOC_CLIENT_RECOVER:
 877                 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
 878                                            data->ioc_inlbuf1, 0);
 879                 if (rc > 0)
 880                         rc = 0;
 881                 break;
 882         case IOC_OSC_SET_ACTIVE:
 883                 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
 884                                               data->ioc_offset);
 885                 break;
 886         case OBD_IOC_PING_TARGET:
 887                 rc = ptlrpc_obd_ping(obd);
 888                 break;
 889         default:
 890                 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
 891                        cmd, cfs_curproc_comm());
 892                 rc = -ENOTTY;
 893         }
 894         cfs_module_put(THIS_MODULE);
 895         return rc;
 896 }
 897
 898 static int osp_obd_health_check(const struct lu_env *env,
 899                                 struct obd_device *obd)
 900 {
 901         struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
 902
 903         ENTRY;
 904
 905         /*
 906          * 1.8/2.0 behaviour is that OST being connected once at least
 907          * is considired "healthy". and one "healty" OST is enough to
 908          * allow lustre clients to connect to MDS
 909          */
 910         LASSERT(d);
 911         RETURN(!d->opd_imp_seen_connected);
 912 }
 913
 914 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
 915 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
 916 static void osp_key_exit(const struct lu_context *ctx,
 917                          struct lu_context_key *key, void *data)
 918 {
 919         struct osp_thread_info *info = data;
 920
 921         info->osi_attr.la_valid = 0;
 922 }
 923
 924 struct lu_context_key osp_thread_key = {
 925         .lct_tags = LCT_MD_THREAD,
 926         .lct_init = osp_key_init,
 927         .lct_fini = osp_key_fini,
 928         .lct_exit = osp_key_exit
 929 };
 930
 931 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
 932 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
 933
 934 struct lu_context_key osp_txn_key = {
 935         .lct_tags = LCT_OSP_THREAD,
 936         .lct_init = osp_txn_key_init,
 937         .lct_fini = osp_txn_key_fini
 938 };
 939 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
 940
 941 static struct lu_device_type_operations osp_device_type_ops = {
 942         .ldto_init           = osp_type_init,
 943         .ldto_fini           = osp_type_fini,
 944
 945         .ldto_start          = osp_type_start,
 946         .ldto_stop           = osp_type_stop,
 947
 948         .ldto_device_alloc   = osp_device_alloc,
 949         .ldto_device_free    = osp_device_free,
 950
 951         .ldto_device_fini    = osp_device_fini
 952 };
 953
 954 static struct lu_device_type osp_device_type = {
 955         .ldt_tags     = LU_DEVICE_DT,
 956         .ldt_name     = LUSTRE_OSP_NAME,
 957         .ldt_ops      = &osp_device_type_ops,
 958         .ldt_ctx_tags = LCT_MD_THREAD
 959 };
 960
 961 static struct obd_ops osp_obd_device_ops = {
 962         .o_owner        = THIS_MODULE,
 963         .o_add_conn     = client_import_add_conn,
 964         .o_del_conn     = client_import_del_conn,
 965         .o_reconnect    = osp_reconnect,
 966         .o_connect      = osp_obd_connect,
 967         .o_disconnect   = osp_obd_disconnect,
 968         .o_health_check = osp_obd_health_check,
 969         .o_import_event = osp_import_event,
 970         .o_iocontrol    = osp_iocontrol,
 971         .o_statfs       = osp_obd_statfs,
 972 };
 973
 974 static int __init osp_mod_init(void)
 975 {
 976         struct lprocfs_static_vars       lvars;
 977         cfs_proc_dir_entry_t            *osc_proc_dir;
 978         int                              rc;
 979
 980         rc = lu_kmem_init(osp_caches);
 981         if (rc)
 982                 return rc;
 983
 984         lprocfs_osp_init_vars(&lvars);
 985
 986         rc = class_register_type(&osp_obd_device_ops, NULL, lvars.module_vars,
 987                                  LUSTRE_OSP_NAME, &osp_device_type);
 988
 989         /* create "osc" entry in procfs for compatibility purposes */
 990         if (rc != 0) {
 991                 lu_kmem_fini(osp_caches);
 992                 return rc;
 993         }
 994
 995         osc_proc_dir = lprocfs_srch(proc_lustre_root, "osc");
 996         if (osc_proc_dir == NULL) {
 997                 osc_proc_dir = lprocfs_register("osc", proc_lustre_root, NULL,
 998                                                 NULL);
 999                 if (IS_ERR(osc_proc_dir))
1000                         CERROR("osp: can't create compat entry \"osc\": %d\n",
1001                                (int) PTR_ERR(osc_proc_dir));
1002         }
1003         return rc;
1004 }
1005
1006 static void __exit osp_mod_exit(void)
1007 {
1008         lprocfs_try_remove_proc_entry("osc", proc_lustre_root);
1009
1010         class_unregister_type(LUSTRE_OSP_NAME);
1011         lu_kmem_fini(osp_caches);
1012 }
1013
1014 MODULE_AUTHOR("Intel, Inc. <http://www.intel.com/>");
1015 MODULE_DESCRIPTION("Lustre OST Proxy Device ("LUSTRE_OSP_NAME")");
1016 MODULE_LICENSE("GPL");
1017
1018 cfs_module(osp, LUSTRE_VERSION_STRING, osp_mod_init, osp_mod_exit);