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 && rc != -ETIMEDOUT)
 216                 CERROR("%s: can't disconnect: rc = %d\n",
 217                        d->opd_obd->obd_name, rc);
 218
 219         ptlrpc_invalidate_import(imp);
 220
 221         RETURN(rc);
 222 }
 223
 224 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
 225 {
 226         int                      rc = 0;
 227         ENTRY;
 228
 229         if (is_osp_on_ost(d->opd_obd->obd_name)) {
 230                 rc = osp_disconnect(d);
 231                 RETURN(rc);
 232         }
 233
 234         LASSERT(env);
 235         /* release last_used file */
 236         osp_last_used_fini(env, d);
 237
 238         rc = osp_disconnect(d);
 239
 240         /* stop precreate thread */
 241         osp_precreate_fini(d);
 242
 243         /* stop sync thread */
 244         osp_sync_fini(d);
 245
 246         RETURN(rc);
 247 }
 248
 249 static int osp_process_config(const struct lu_env *env,
 250                               struct lu_device *dev, struct lustre_cfg *lcfg)
 251 {
 252         struct osp_device               *d = lu2osp_dev(dev);
 253         struct lprocfs_static_vars       lvars = { 0 };
 254         int                              rc;
 255
 256         ENTRY;
 257
 258         switch (lcfg->lcfg_command) {
 259         case LCFG_CLEANUP:
 260                 if (!is_osp_on_ost(d->opd_obd->obd_name))
 261                         lu_dev_del_linkage(dev->ld_site, dev);
 262                 rc = osp_shutdown(env, d);
 263                 break;
 264         case LCFG_PARAM:
 265                 lprocfs_osp_init_vars(&lvars);
 266
 267                 LASSERT(d->opd_obd);
 268                 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
 269                                               lcfg, d->opd_obd);
 270                 if (rc > 0)
 271                         rc = 0;
 272                 if (rc == -ENOSYS) {
 273                         /* class_process_proc_param() haven't found matching
 274                          * parameter and returned ENOSYS so that layer(s)
 275                          * below could use that. But OSP is the bottom, so
 276                          * just ignore it */
 277                         CERROR("%s: unknown param %s\n",
 278                                (char *)lustre_cfg_string(lcfg, 0),
 279                                (char *)lustre_cfg_string(lcfg, 1));
 280                         rc = 0;
 281                 }
 282                 break;
 283         default:
 284                 CERROR("%s: unknown command %u\n",
 285                        (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
 286                 rc = 0;
 287                 break;
 288         }
 289
 290         RETURN(rc);
 291 }
 292
 293 static int osp_recovery_complete(const struct lu_env *env,
 294                                  struct lu_device *dev)
 295 {
 296         struct osp_device       *osp = lu2osp_dev(dev);
 297         int                      rc = 0;
 298
 299         ENTRY;
 300         osp->opd_recovery_completed = 1;
 301         cfs_waitq_signal(&osp->opd_pre_waitq);
 302         RETURN(rc);
 303 }
 304
 305 const struct lu_device_operations osp_lu_ops = {
 306         .ldo_object_alloc       = osp_object_alloc,
 307         .ldo_process_config     = osp_process_config,
 308         .ldo_recovery_complete  = osp_recovery_complete,
 309 };
 310
 311 /**
 312  * provides with statfs from corresponded OST
 313  *
 314  */
 315 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
 316                       struct obd_statfs *sfs)
 317 {
 318         struct osp_device *d = dt2osp_dev(dev);
 319
 320         ENTRY;
 321
 322         if (unlikely(d->opd_imp_active == 0))
 323                 RETURN(-ENOTCONN);
 324
 325         /* return recently updated data */
 326         *sfs = d->opd_statfs;
 327
 328         /*
 329          * layer above osp (usually lod) can use ffree to estimate
 330          * how many objects are available for immediate creation
 331          */
 332         cfs_spin_lock(&d->opd_pre_lock);
 333         sfs->os_fprecreated = d->opd_pre_last_created - d->opd_pre_next;
 334         cfs_spin_unlock(&d->opd_pre_lock);
 335
 336         CDEBUG(D_OTHER, "%s: "LPU64" blocks, "LPU64" free, "LPU64" avail, "
 337                LPU64" files, "LPU64" free files\n", d->opd_obd->obd_name,
 338                sfs->os_blocks, sfs->os_bfree, sfs->os_bavail,
 339                sfs->os_files, sfs->os_ffree);
 340         RETURN(0);
 341 }
 342
 343 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
 344 {
 345         ENTRY;
 346
 347         /*
 348          * XXX: wake up sync thread, command it to start flushing asap?
 349          */
 350
 351         RETURN(0);
 352 }
 353
 354 const struct dt_device_operations osp_dt_ops = {
 355         .dt_statfs      = osp_statfs,
 356         .dt_sync        = osp_sync,
 357 };
 358
 359 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *m,
 360                               const char *nextdev)
 361 {
 362         struct obd_connect_data *data = NULL;
 363         struct obd_device       *obd;
 364         int                      rc;
 365
 366         ENTRY;
 367
 368         LASSERT(m->opd_storage_exp == NULL);
 369
 370         OBD_ALLOC_PTR(data);
 371         if (data == NULL)
 372                 RETURN(-ENOMEM);
 373
 374         obd = class_name2obd(nextdev);
 375         if (obd == NULL) {
 376                 CERROR("%s: can't locate next device: %s\n",
 377                        m->opd_obd->obd_name, nextdev);
 378                 GOTO(out, rc = -ENOTCONN);
 379         }
 380
 381         rc = obd_connect(env, &m->opd_storage_exp, obd, &obd->obd_uuid, data,
 382                          NULL);
 383         if (rc) {
 384                 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
 385                        m->opd_obd->obd_name, nextdev, rc);
 386                 GOTO(out, rc);
 387         }
 388
 389         m->opd_dt_dev.dd_lu_dev.ld_site =
 390                 m->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
 391         LASSERT(m->opd_dt_dev.dd_lu_dev.ld_site);
 392         m->opd_storage = lu2dt_dev(m->opd_storage_exp->exp_obd->obd_lu_dev);
 393
 394 out:
 395         OBD_FREE_PTR(data);
 396         RETURN(rc);
 397 }
 398
 399 static int osp_init0(const struct lu_env *env, struct osp_device *m,
 400                      struct lu_device_type *ldt, struct lustre_cfg *cfg)
 401 {
 402         struct lprocfs_static_vars       lvars = { 0 };
 403         struct proc_dir_entry           *osc_proc_dir;
 404         struct obd_import               *imp;
 405         class_uuid_t                     uuid;
 406         char                            *src, *ost, *mdt, *osdname = NULL;
 407         int                              rc, idx;
 408
 409         ENTRY;
 410
 411         m->opd_obd = class_name2obd(lustre_cfg_string(cfg, 0));
 412         if (m->opd_obd == NULL) {
 413                 CERROR("Cannot find obd with name %s\n",
 414                        lustre_cfg_string(cfg, 0));
 415                 RETURN(-ENODEV);
 416         }
 417
 418         /* There is no record in the MDT configuration for the local disk
 419          * device, so we have to extract this from elsewhere in the profile.
 420          * The only information we get at setup is from the OSC records:
 421          * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 422          * Note that 1.8 generated configs are missing the -MDTxxxx part.
 423          * We need to reconstruct the name of the underlying OSD from this:
 424          * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".  We
 425          * also need to determine the OST index from this - will be used
 426          * to calculate the offset in shared lov_objids file later */
 427
 428         src = lustre_cfg_string(cfg, 0);
 429         if (src == NULL)
 430                 RETURN(-EINVAL);
 431
 432         ost = strstr(src, "-OST");
 433         if (ost == NULL)
 434                 RETURN(-EINVAL);
 435
 436         idx = simple_strtol(ost + 4, &mdt, 16);
 437         if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
 438                 CERROR("%s: invalid OST index in '%s'\n",
 439                        m->opd_obd->obd_name, src);
 440                 GOTO(out_fini, rc = -EINVAL);
 441         }
 442         m->opd_index = idx;
 443
 444         idx = ost - src;
 445         /* check the fsname length, and after this everything else will fit */
 446         if (idx > MTI_NAME_MAXLEN) {
 447                 CERROR("%s: fsname too long in '%s'\n",
 448                        m->opd_obd->obd_name, src);
 449                 GOTO(out_fini, rc = -EINVAL);
 450         }
 451
 452         OBD_ALLOC(osdname, MAX_OBD_NAME);
 453         if (osdname == NULL)
 454                 GOTO(out_fini, rc = -ENOMEM);
 455
 456         memcpy(osdname, src, idx); /* copy just the fsname part */
 457         osdname[idx] = '\0';
 458
 459         mdt = strstr(mdt, "-MDT");
 460         if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
 461                 strcat(osdname, "-MDT0000");
 462         else
 463                 strcat(osdname, mdt);
 464         strcat(osdname, "-osd");
 465         CDEBUG(D_HA, "%s: connect to %s (%s)\n",
 466                m->opd_obd->obd_name, osdname, src);
 467
 468         m->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
 469         m->opd_dt_dev.dd_ops = &osp_dt_ops;
 470         m->opd_obd->obd_lu_dev = &m->opd_dt_dev.dd_lu_dev;
 471
 472         rc = osp_connect_to_osd(env, m, osdname);
 473         if (rc)
 474                 GOTO(out_fini, rc);
 475
 476         rc = ptlrpcd_addref();
 477         if (rc)
 478                 GOTO(out_disconnect, rc);
 479
 480         rc = client_obd_setup(m->opd_obd, cfg);
 481         if (rc) {
 482                 CERROR("%s: can't setup obd: %d\n", m->opd_obd->obd_name, rc);
 483                 GOTO(out_ref, rc);
 484         }
 485
 486         lprocfs_osp_init_vars(&lvars);
 487         if (lprocfs_obd_setup(m->opd_obd, lvars.obd_vars) == 0)
 488                 ptlrpc_lprocfs_register_obd(m->opd_obd);
 489
 490         /* for compatibility we link old procfs's OSC entries to osp ones */
 491         osc_proc_dir = lprocfs_srch(proc_lustre_root, "osc");
 492         if (osc_proc_dir) {
 493                 cfs_proc_dir_entry_t    *symlink = NULL;
 494                 char                    *name;
 495
 496                 OBD_ALLOC(name, strlen(m->opd_obd->obd_name) + 1);
 497                 if (name == NULL)
 498                         GOTO(out, rc = -ENOMEM);
 499
 500                 strcpy(name, m->opd_obd->obd_name);
 501                 if (strstr(name, "osc"))
 502                         symlink = lprocfs_add_symlink(name, osc_proc_dir,
 503                                                       "../osp/%s",
 504                                                       m->opd_obd->obd_name);
 505                 OBD_FREE(name, strlen(m->opd_obd->obd_name) + 1);
 506                 m->opd_symlink = symlink;
 507         }
 508
 509         /*
 510          * Initialize last id from the storage - will be used in orphan cleanup
 511          */
 512         rc = osp_last_used_init(env, m);
 513         if (rc)
 514                 GOTO(out_proc, rc);
 515
 516         /*
 517          * Initialize precreation thread, it handles new connections as well
 518          */
 519         rc = osp_init_precreate(m);
 520         if (rc)
 521                 GOTO(out_last_used, rc);
 522
 523         /*
 524          * Initialize synhronization mechanism taking care of propogating
 525          * changes to OST in near transactional manner
 526          */
 527         rc = osp_sync_init(env, m);
 528         if (rc)
 529                 GOTO(out_precreat, rc);
 530
 531         /*
 532          * Initiate connect to OST
 533          */
 534         ll_generate_random_uuid(uuid);
 535         class_uuid_unparse(uuid, &m->opd_cluuid);
 536
 537         imp = m->opd_obd->u.cli.cl_import;
 538
 539         rc = ptlrpc_init_import(imp);
 540         if (rc)
 541                 GOTO(out, rc);
 542         if (osdname)
 543                 OBD_FREE(osdname, MAX_OBD_NAME);
 544         RETURN(0);
 545
 546 out:
 547         /* stop sync thread */
 548         osp_sync_fini(m);
 549 out_precreat:
 550         /* stop precreate thread */
 551         osp_precreate_fini(m);
 552 out_last_used:
 553         osp_last_used_fini(env, m);
 554 out_proc:
 555         ptlrpc_lprocfs_unregister_obd(m->opd_obd);
 556         lprocfs_obd_cleanup(m->opd_obd);
 557         class_destroy_import(m->opd_obd->u.cli.cl_import);
 558         client_obd_cleanup(m->opd_obd);
 559 out_ref:
 560         ptlrpcd_decref();
 561 out_disconnect:
 562         obd_disconnect(m->opd_storage_exp);
 563 out_fini:
 564         if (osdname)
 565                 OBD_FREE(osdname, MAX_OBD_NAME);
 566         RETURN(rc);
 567 }
 568
 569 static struct lu_device *osp_device_free(const struct lu_env *env,
 570                                          struct lu_device *lu)
 571 {
 572         struct osp_device *m = lu2osp_dev(lu);
 573
 574         ENTRY;
 575
 576         if (cfs_atomic_read(&lu->ld_ref) && lu->ld_site) {
 577                 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
 578                 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
 579         }
 580         dt_device_fini(&m->opd_dt_dev);
 581         OBD_FREE_PTR(m);
 582         RETURN(NULL);
 583 }
 584
 585 static struct lu_device *osp_device_alloc(const struct lu_env *env,
 586                                           struct lu_device_type *t,
 587                                           struct lustre_cfg *lcfg)
 588 {
 589         struct osp_device *m;
 590         struct lu_device  *l;
 591
 592         OBD_ALLOC_PTR(m);
 593         if (m == NULL) {
 594                 l = ERR_PTR(-ENOMEM);
 595         } else {
 596                 int rc;
 597
 598                 l = osp2lu_dev(m);
 599                 dt_device_init(&m->opd_dt_dev, t);
 600                 if (is_osp_on_ost(lustre_cfg_string(lcfg, 0)))
 601                         rc = osp_init_for_ost(env, m, t, lcfg);
 602                 else
 603                         rc = osp_init0(env, m, t, lcfg);
 604                 if (rc != 0) {
 605                         osp_device_free(env, l);
 606                         l = ERR_PTR(rc);
 607                 }
 608         }
 609         return l;
 610 }
 611
 612 static struct lu_device *osp_device_fini(const struct lu_env *env,
 613                                          struct lu_device *d)
 614 {
 615         struct osp_device *m = lu2osp_dev(d);
 616         struct obd_import *imp;
 617         int                rc;
 618
 619         ENTRY;
 620
 621         if (m->opd_storage_exp)
 622                 obd_disconnect(m->opd_storage_exp);
 623
 624         if (is_osp_on_ost(m->opd_obd->obd_name))
 625                 osp_fini_for_ost(m);
 626
 627         imp = m->opd_obd->u.cli.cl_import;
 628
 629         if (imp->imp_rq_pool) {
 630                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 631                 imp->imp_rq_pool = NULL;
 632         }
 633
 634         obd_cleanup_client_import(m->opd_obd);
 635
 636         if (m->opd_symlink)
 637                 lprocfs_remove(&m->opd_symlink);
 638
 639         LASSERT(m->opd_obd);
 640         ptlrpc_lprocfs_unregister_obd(m->opd_obd);
 641         lprocfs_obd_cleanup(m->opd_obd);
 642
 643         rc = client_obd_cleanup(m->opd_obd);
 644         LASSERTF(rc == 0, "error %d\n", rc);
 645
 646         ptlrpcd_decref();
 647
 648         RETURN(NULL);
 649 }
 650
 651 static int osp_reconnect(const struct lu_env *env,
 652                          struct obd_export *exp, struct obd_device *obd,
 653                          struct obd_uuid *cluuid,
 654                          struct obd_connect_data *data,
 655                          void *localdata)
 656 {
 657         return 0;
 658 }
 659
 660 /*
 661  * we use exports to track all LOD users
 662  */
 663 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
 664                            struct obd_device *obd, struct obd_uuid *cluuid,
 665                            struct obd_connect_data *data, void *localdata)
 666 {
 667         struct osp_device       *osp = lu2osp_dev(obd->obd_lu_dev);
 668         struct obd_connect_data *ocd;
 669         struct obd_import       *imp;
 670         struct lustre_handle     conn;
 671         int                      rc;
 672
 673         ENTRY;
 674
 675         CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
 676
 677         rc = class_connect(&conn, obd, cluuid);
 678         if (rc)
 679                 RETURN(rc);
 680
 681         *exp = class_conn2export(&conn);
 682         if (is_osp_on_ost(obd->obd_name))
 683                 osp->opd_exp = *exp;
 684
 685         /* Why should there ever be more than 1 connect? */
 686         osp->opd_connects++;
 687         LASSERT(osp->opd_connects == 1);
 688
 689         imp = osp->opd_obd->u.cli.cl_import;
 690         imp->imp_dlm_handle = conn;
 691
 692         ocd = &imp->imp_connect_data;
 693         ocd->ocd_connect_flags = OBD_CONNECT_AT |
 694                                  OBD_CONNECT_FULL20 |
 695                                  OBD_CONNECT_INDEX |
 696 #ifdef HAVE_LRU_RESIZE_SUPPORT
 697                                  OBD_CONNECT_LRU_RESIZE |
 698 #endif
 699                                  OBD_CONNECT_MDS |
 700                                  OBD_CONNECT_OSS_CAPA |
 701                                  OBD_CONNECT_REQPORTAL |
 702                                  OBD_CONNECT_SKIP_ORPHAN |
 703                                  OBD_CONNECT_VERSION |
 704                                  OBD_CONNECT_FID;
 705
 706         if (is_osp_on_ost(osp->opd_obd->obd_name))
 707                 ocd->ocd_connect_flags |= OBD_CONNECT_LIGHTWEIGHT;
 708
 709         ocd->ocd_version = LUSTRE_VERSION_CODE;
 710         LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
 711         ocd->ocd_index = data->ocd_index;
 712         imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
 713
 714         rc = ptlrpc_connect_import(imp);
 715         if (rc) {
 716                 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
 717                 GOTO(out, rc);
 718         }
 719
 720         ptlrpc_pinger_add_import(imp);
 721
 722 out:
 723         RETURN(rc);
 724 }
 725
 726 /*
 727  * once last export (we don't count self-export) disappeared
 728  * osp can be released
 729  */
 730 static int osp_obd_disconnect(struct obd_export *exp)
 731 {
 732         struct obd_device *obd = exp->exp_obd;
 733         struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
 734         int                rc;
 735         ENTRY;
 736
 737         /* Only disconnect the underlying layers on the final disconnect. */
 738         LASSERT(osp->opd_connects == 1);
 739         osp->opd_connects--;
 740
 741         rc = class_disconnect(exp);
 742         if (rc) {
 743                 CERROR("%s: class disconnect error: rc = %d\n",
 744                        obd->obd_name, rc);
 745                 RETURN(rc);
 746         }
 747
 748         /* destroy the device */
 749         if (!is_osp_on_ost(obd->obd_name))
 750                 class_manual_cleanup(obd);
 751
 752         RETURN(rc);
 753 }
 754
 755 /*
 756  * lprocfs helpers still use OBD API, let's keep obd_statfs() support
 757  */
 758 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
 759                           struct obd_statfs *osfs, __u64 max_age, __u32 flags)
 760 {
 761         struct obd_statfs       *msfs;
 762         struct ptlrpc_request   *req;
 763         struct obd_import       *imp = NULL;
 764         int                      rc;
 765
 766         ENTRY;
 767
 768         /* Since the request might also come from lprocfs, so we need
 769          * sync this with client_disconnect_export Bug15684 */
 770         cfs_down_read(&exp->exp_obd->u.cli.cl_sem);
 771         if (exp->exp_obd->u.cli.cl_import)
 772                 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
 773         cfs_up_read(&exp->exp_obd->u.cli.cl_sem);
 774         if (!imp)
 775                 RETURN(-ENODEV);
 776
 777         /* We could possibly pass max_age in the request (as an absolute
 778          * timestamp or a "seconds.usec ago") so the target can avoid doing
 779          * extra calls into the filesystem if that isn't necessary (e.g.
 780          * during mount that would help a bit).  Having relative timestamps
 781          * is not so great if request processing is slow, while absolute
 782          * timestamps are not ideal because they need time synchronization. */
 783         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
 784
 785         class_import_put(imp);
 786
 787         if (req == NULL)
 788                 RETURN(-ENOMEM);
 789
 790         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
 791         if (rc) {
 792                 ptlrpc_request_free(req);
 793                 RETURN(rc);
 794         }
 795         ptlrpc_request_set_replen(req);
 796         req->rq_request_portal = OST_CREATE_PORTAL;
 797         ptlrpc_at_set_req_timeout(req);
 798
 799         if (flags & OBD_STATFS_NODELAY) {
 800                 /* procfs requests not want stat in wait for avoid deadlock */
 801                 req->rq_no_resend = 1;
 802                 req->rq_no_delay = 1;
 803         }
 804
 805         rc = ptlrpc_queue_wait(req);
 806         if (rc)
 807                 GOTO(out, rc);
 808
 809         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
 810         if (msfs == NULL)
 811                 GOTO(out, rc = -EPROTO);
 812
 813         *osfs = *msfs;
 814
 815         EXIT;
 816 out:
 817         ptlrpc_req_finished(req);
 818         return rc;
 819 }
 820
 821 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
 822                             enum obd_import_event event)
 823 {
 824         struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
 825
 826         switch (event) {
 827         case IMP_EVENT_DISCON:
 828                 d->opd_got_disconnected = 1;
 829                 d->opd_imp_connected = 0;
 830                 if (is_osp_on_ost(d->opd_obd->obd_name))
 831                         break;
 832                 osp_pre_update_status(d, -ENODEV);
 833                 cfs_waitq_signal(&d->opd_pre_waitq);
 834                 CDEBUG(D_HA, "got disconnected\n");
 835                 break;
 836         case IMP_EVENT_INACTIVE:
 837                 d->opd_imp_active = 0;
 838                 if (is_osp_on_ost(d->opd_obd->obd_name))
 839                         break;
 840                 osp_pre_update_status(d, -ENODEV);
 841                 cfs_waitq_signal(&d->opd_pre_waitq);
 842                 CDEBUG(D_HA, "got inactive\n");
 843                 break;
 844         case IMP_EVENT_ACTIVE:
 845                 d->opd_imp_active = 1;
 846                 if (d->opd_got_disconnected)
 847                         d->opd_new_connection = 1;
 848                 d->opd_imp_connected = 1;
 849                 d->opd_imp_seen_connected = 1;
 850                 if (is_osp_on_ost(d->opd_obd->obd_name))
 851                         break;
 852                 cfs_waitq_signal(&d->opd_pre_waitq);
 853                 __osp_sync_check_for_work(d);
 854                 CDEBUG(D_HA, "got connected\n");
 855                 break;
 856         case IMP_EVENT_INVALIDATE:
 857                 if (obd->obd_namespace == NULL)
 858                         break;
 859                 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
 860                 break;
 861         case IMP_EVENT_OCD:
 862                 break;
 863         default:
 864                 CERROR("%s: unsupported import event: %#x\n",
 865                        obd->obd_name, event);
 866         }
 867         return 0;
 868 }
 869
 870 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
 871                          void *karg, void *uarg)
 872 {
 873         struct obd_device       *obd = exp->exp_obd;
 874         struct osp_device       *d;
 875         struct obd_ioctl_data   *data = karg;
 876         int                      rc = 0;
 877
 878         ENTRY;
 879
 880         LASSERT(obd->obd_lu_dev);
 881         d = lu2osp_dev(obd->obd_lu_dev);
 882         LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
 883
 884         if (!cfs_try_module_get(THIS_MODULE)) {
 885                 CERROR("%s: can't get module. Is it alive?", obd->obd_name);
 886                 return -EINVAL;
 887         }
 888
 889         switch (cmd) {
 890         case OBD_IOC_CLIENT_RECOVER:
 891                 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
 892                                            data->ioc_inlbuf1, 0);
 893                 if (rc > 0)
 894                         rc = 0;
 895                 break;
 896         case IOC_OSC_SET_ACTIVE:
 897                 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
 898                                               data->ioc_offset);
 899                 break;
 900         case OBD_IOC_PING_TARGET:
 901                 rc = ptlrpc_obd_ping(obd);
 902                 break;
 903         default:
 904                 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
 905                        cmd, cfs_curproc_comm());
 906                 rc = -ENOTTY;
 907         }
 908         cfs_module_put(THIS_MODULE);
 909         return rc;
 910 }
 911
 912 static int osp_obd_health_check(const struct lu_env *env,
 913                                 struct obd_device *obd)
 914 {
 915         struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
 916
 917         ENTRY;
 918
 919         /*
 920          * 1.8/2.0 behaviour is that OST being connected once at least
 921          * is considired "healthy". and one "healty" OST is enough to
 922          * allow lustre clients to connect to MDS
 923          */
 924         LASSERT(d);
 925         RETURN(!d->opd_imp_seen_connected);
 926 }
 927
 928 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
 929 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
 930 static void osp_key_exit(const struct lu_context *ctx,
 931                          struct lu_context_key *key, void *data)
 932 {
 933         struct osp_thread_info *info = data;
 934
 935         info->osi_attr.la_valid = 0;
 936 }
 937
 938 struct lu_context_key osp_thread_key = {
 939         .lct_tags = LCT_MD_THREAD,
 940         .lct_init = osp_key_init,
 941         .lct_fini = osp_key_fini,
 942         .lct_exit = osp_key_exit
 943 };
 944
 945 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
 946 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
 947
 948 struct lu_context_key osp_txn_key = {
 949         .lct_tags = LCT_OSP_THREAD,
 950         .lct_init = osp_txn_key_init,
 951         .lct_fini = osp_txn_key_fini
 952 };
 953 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
 954
 955 static struct lu_device_type_operations osp_device_type_ops = {
 956         .ldto_init           = osp_type_init,
 957         .ldto_fini           = osp_type_fini,
 958
 959         .ldto_start          = osp_type_start,
 960         .ldto_stop           = osp_type_stop,
 961
 962         .ldto_device_alloc   = osp_device_alloc,
 963         .ldto_device_free    = osp_device_free,
 964
 965         .ldto_device_fini    = osp_device_fini
 966 };
 967
 968 static struct lu_device_type osp_device_type = {
 969         .ldt_tags     = LU_DEVICE_DT,
 970         .ldt_name     = LUSTRE_OSP_NAME,
 971         .ldt_ops      = &osp_device_type_ops,
 972         .ldt_ctx_tags = LCT_MD_THREAD
 973 };
 974
 975 static struct obd_ops osp_obd_device_ops = {
 976         .o_owner        = THIS_MODULE,
 977         .o_add_conn     = client_import_add_conn,
 978         .o_del_conn     = client_import_del_conn,
 979         .o_reconnect    = osp_reconnect,
 980         .o_connect      = osp_obd_connect,
 981         .o_disconnect   = osp_obd_disconnect,
 982         .o_health_check = osp_obd_health_check,
 983         .o_import_event = osp_import_event,
 984         .o_iocontrol    = osp_iocontrol,
 985         .o_statfs       = osp_obd_statfs,
 986 };
 987
 988 static int __init osp_mod_init(void)
 989 {
 990         struct lprocfs_static_vars       lvars;
 991         cfs_proc_dir_entry_t            *osc_proc_dir;
 992         int                              rc;
 993
 994         rc = lu_kmem_init(osp_caches);
 995         if (rc)
 996                 return rc;
 997
 998         lprocfs_osp_init_vars(&lvars);
 999
1000         rc = class_register_type(&osp_obd_device_ops, NULL, lvars.module_vars,
1001                                  LUSTRE_OSP_NAME, &osp_device_type);
1002
1003         /* create "osc" entry in procfs for compatibility purposes */
1004         if (rc != 0) {
1005                 lu_kmem_fini(osp_caches);
1006                 return rc;
1007         }
1008
1009         osc_proc_dir = lprocfs_srch(proc_lustre_root, "osc");
1010         if (osc_proc_dir == NULL) {
1011                 osc_proc_dir = lprocfs_register("osc", proc_lustre_root, NULL,
1012                                                 NULL);
1013                 if (IS_ERR(osc_proc_dir))
1014                         CERROR("osp: can't create compat entry \"osc\": %d\n",
1015                                (int) PTR_ERR(osc_proc_dir));
1016         }
1017         return rc;
1018 }
1019
1020 static void __exit osp_mod_exit(void)
1021 {
1022         lprocfs_try_remove_proc_entry("osc", proc_lustre_root);
1023
1024         class_unregister_type(LUSTRE_OSP_NAME);
1025         lu_kmem_fini(osp_caches);
1026 }
1027
1028 MODULE_AUTHOR("Intel, Inc. <http://www.intel.com/>");
1029 MODULE_DESCRIPTION("Lustre OST Proxy Device ("LUSTRE_OSP_NAME")");
1030 MODULE_LICENSE("GPL");
1031
1032 cfs_module(osp, LUSTRE_VERSION_STRING, osp_mod_init, osp_mod_exit);
1033