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