lustre/osp/osp_dev.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/osp/osp_dev.c
  37  *
  38  * Lustre OST Proxy Device
  39  *
  40  * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
  41  * Author: Mikhail Pershin <mike.pershin@intel.com>
  42  * Author: Di Wang <di.wang@intel.com>
  43  */
  44
  45 #ifndef EXPORT_SYMTAB
  46 # define EXPORT_SYMTAB
  47 #endif
  48 #define DEBUG_SUBSYSTEM S_MDS
  49
  50 #include <obd_class.h>
  51 #include <lustre_param.h>
  52 #include <lustre_log.h>
  53 #include <lustre_mdc.h>
  54
  55 #include "osp_internal.h"
  56
  57 /* Slab for OSP object allocation */
  58 cfs_mem_cache_t *osp_object_kmem;
  59
  60 static struct lu_kmem_descr osp_caches[] = {
  61         {
  62                 .ckd_cache = &osp_object_kmem,
  63                 .ckd_name  = "osp_obj",
  64                 .ckd_size  = sizeof(struct osp_object)
  65         },
  66         {
  67                 .ckd_cache = NULL
  68         }
  69 };
  70
  71 struct lu_object *osp_object_alloc(const struct lu_env *env,
  72                                    const struct lu_object_header *hdr,
  73                                    struct lu_device *d)
  74 {
  75         struct lu_object_header *h = NULL;
  76         struct osp_object       *o;
  77         struct lu_object        *l;
  78
  79         OBD_SLAB_ALLOC_PTR_GFP(o, osp_object_kmem, CFS_ALLOC_IO);
  80         if (o != NULL) {
  81                 l = &o->opo_obj.do_lu;
  82
  83                 /* For data object, OSP obj would always be the top
  84                  * object, i.e. hdr is always NULL, see lu_object_alloc.
  85                  * But for metadata object, we always build the object
  86                  * stack from MDT. i.e. mdt_object will be the top object
  87                  * i.e.  hdr != NULL */
  88                 if (hdr == NULL) {
  89                         /* object for OST */
  90                         h = &o->opo_header;
  91                         lu_object_header_init(h);
  92                         dt_object_init(&o->opo_obj, h, d);
  93                         lu_object_add_top(h, l);
  94                 } else {
  95                         dt_object_init(&o->opo_obj, h, d);
  96                 }
  97
  98                 l->lo_ops = &osp_lu_obj_ops;
  99
 100                 return l;
 101         } else {
 102                 return NULL;
 103         }
 104 }
 105
 106 static struct dt_object
 107 *osp_find_or_create(const struct lu_env *env, struct osp_device *osp,
 108                     struct lu_attr *attr, __u32 reg_id)
 109 {
 110         struct osp_thread_info *osi = osp_env_info(env);
 111         struct dt_object_format dof = { 0 };
 112         struct dt_object       *dto;
 113         int                  rc;
 114         ENTRY;
 115
 116         lu_local_obj_fid(&osi->osi_fid, reg_id);
 117         attr->la_valid = LA_MODE;
 118         attr->la_mode = S_IFREG | 0644;
 119         dof.dof_type = DFT_REGULAR;
 120         dto = dt_find_or_create(env, osp->opd_storage, &osi->osi_fid,
 121                                 &dof, attr);
 122         if (IS_ERR(dto))
 123                 RETURN(dto);
 124
 125         rc = dt_attr_get(env, dto, attr, NULL);
 126         if (rc) {
 127                 CERROR("%s: can't be initialized: rc = %d\n",
 128                        osp->opd_obd->obd_name, rc);
 129                 lu_object_put(env, &dto->do_lu);
 130                 RETURN(ERR_PTR(rc));
 131         }
 132         RETURN(dto);
 133 }
 134
 135 static int osp_write_local_file(const struct lu_env *env,
 136                                 struct osp_device *osp,
 137                                 struct dt_object *dt_obj,
 138                                 struct lu_buf *buf,
 139                                 loff_t offset)
 140 {
 141         struct thandle *th;
 142         int rc;
 143
 144         th = dt_trans_create(env, osp->opd_storage);
 145         if (IS_ERR(th))
 146                 RETURN(PTR_ERR(th));
 147
 148         rc = dt_declare_record_write(env, dt_obj, buf->lb_len, offset, th);
 149         if (rc)
 150                 GOTO(out, rc);
 151         rc = dt_trans_start_local(env, osp->opd_storage, th);
 152         if (rc)
 153                 GOTO(out, rc);
 154
 155         rc = dt_record_write(env, dt_obj, buf, &offset, th);
 156 out:
 157         dt_trans_stop(env, osp->opd_storage, th);
 158         RETURN(rc);
 159 }
 160
 161 static int osp_init_last_objid(const struct lu_env *env, struct osp_device *osp)
 162 {
 163         struct osp_thread_info  *osi = osp_env_info(env);
 164         struct lu_fid           *fid = &osp->opd_last_used_fid;
 165         struct dt_object        *dto;
 166         int                     rc;
 167         ENTRY;
 168
 169         dto = osp_find_or_create(env, osp, &osi->osi_attr, MDD_LOV_OBJ_OID);
 170         if (IS_ERR(dto))
 171                 RETURN(PTR_ERR(dto));
 172         /* object will be released in device cleanup path */
 173         if (osi->osi_attr.la_size >=
 174             sizeof(osi->osi_id) * (osp->opd_index + 1)) {
 175                 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
 176                                    osp->opd_index);
 177                 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
 178                 if (rc != 0)
 179                         GOTO(out, rc);
 180         } else {
 181                 fid->f_oid = 0;
 182                 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
 183                                    osp->opd_index);
 184                 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
 185                                           osi->osi_off);
 186         }
 187         osp->opd_last_used_oid_file = dto;
 188         RETURN(0);
 189 out:
 190         /* object will be released in device cleanup path */
 191         CERROR("%s: can't initialize lov_objid: rc = %d\n",
 192                osp->opd_obd->obd_name, rc);
 193         lu_object_put(env, &dto->do_lu);
 194         osp->opd_last_used_oid_file = NULL;
 195         RETURN(rc);
 196 }
 197
 198 static int osp_init_last_seq(const struct lu_env *env, struct osp_device *osp)
 199 {
 200         struct osp_thread_info  *osi = osp_env_info(env);
 201         struct lu_fid           *fid = &osp->opd_last_used_fid;
 202         struct dt_object        *dto;
 203         int                     rc;
 204         ENTRY;
 205
 206         dto = osp_find_or_create(env, osp, &osi->osi_attr, MDD_LOV_OBJ_OSEQ);
 207         if (IS_ERR(dto))
 208                 RETURN(PTR_ERR(dto));
 209
 210         /* object will be released in device cleanup path */
 211         if (osi->osi_attr.la_size >=
 212             sizeof(osi->osi_id) * (osp->opd_index + 1)) {
 213                 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
 214                                    osp->opd_index);
 215                 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
 216                 if (rc != 0)
 217                         GOTO(out, rc);
 218         } else {
 219                 fid->f_seq = 0;
 220                 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
 221                                     osp->opd_index);
 222                 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
 223                                           osi->osi_off);
 224         }
 225         osp->opd_last_used_seq_file = dto;
 226         RETURN(0);
 227 out:
 228         /* object will be released in device cleanup path */
 229         CERROR("%s: can't initialize lov_seq: rc = %d\n",
 230                osp->opd_obd->obd_name, rc);
 231         lu_object_put(env, &dto->do_lu);
 232         osp->opd_last_used_seq_file = NULL;
 233         RETURN(rc);
 234 }
 235
 236 static int osp_last_used_init(const struct lu_env *env, struct osp_device *osp)
 237 {
 238         struct osp_thread_info *osi = osp_env_info(env);
 239         int                  rc;
 240         ENTRY;
 241
 242         fid_zero(&osp->opd_last_used_fid);
 243         rc = osp_init_last_objid(env, osp);
 244         if (rc < 0) {
 245                 CERROR("%s: Can not get ids %d from old objid!\n",
 246                        osp->opd_obd->obd_name, rc);
 247                 RETURN(rc);
 248         }
 249
 250         rc = osp_init_last_seq(env, osp);
 251         if (rc < 0) {
 252                 CERROR("%s: Can not get ids %d from old objid!\n",
 253                        osp->opd_obd->obd_name, rc);
 254                 GOTO(out, rc);
 255         }
 256
 257         if (fid_oid(&osp->opd_last_used_fid) != 0 &&
 258             fid_seq(&osp->opd_last_used_fid) == 0) {
 259                 /* Just upgrade from the old version,
 260                  * set the seq to be IDIF */
 261                 osp->opd_last_used_fid.f_seq =
 262                    fid_idif_seq(fid_oid(&osp->opd_last_used_fid),
 263                                 osp->opd_index);
 264                 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off,
 265                                     &osp->opd_last_used_fid.f_seq,
 266                                     osp->opd_index);
 267                 rc = osp_write_local_file(env, osp, osp->opd_last_used_seq_file,
 268                                           &osi->osi_lb, osi->osi_off);
 269                 if (rc) {
 270                         CERROR("%s : Can not write seq file: rc = %d\n",
 271                                osp->opd_obd->obd_name, rc);
 272                         GOTO(out, rc);
 273                 }
 274         }
 275
 276         if (!fid_is_zero(&osp->opd_last_used_fid) &&
 277                  !fid_is_sane(&osp->opd_last_used_fid)) {
 278                 CERROR("%s: Got invalid FID "DFID"\n", osp->opd_obd->obd_name,
 279                         PFID(&osp->opd_last_used_fid));
 280                 GOTO(out, rc = -EINVAL);
 281         }
 282
 283         CDEBUG(D_INFO, "%s: Init last used fid "DFID"\n",
 284                osp->opd_obd->obd_name, PFID(&osp->opd_last_used_fid));
 285 out:
 286         if (rc != 0) {
 287                 if (osp->opd_last_used_oid_file != NULL) {
 288                         lu_object_put(env, &osp->opd_last_used_oid_file->do_lu);
 289                         osp->opd_last_used_oid_file = NULL;
 290                 }
 291                 if (osp->opd_last_used_seq_file != NULL) {
 292                         lu_object_put(env, &osp->opd_last_used_seq_file->do_lu);
 293                         osp->opd_last_used_seq_file = NULL;
 294                 }
 295         }
 296
 297         RETURN(rc);
 298 }
 299
 300 static void osp_last_used_fini(const struct lu_env *env, struct osp_device *d)
 301 {
 302         /* release last_used file */
 303         if (d->opd_last_used_oid_file != NULL) {
 304                 lu_object_put(env, &d->opd_last_used_oid_file->do_lu);
 305                 d->opd_last_used_oid_file = NULL;
 306         }
 307
 308         if (d->opd_last_used_seq_file != NULL) {
 309                 lu_object_put(env, &d->opd_last_used_seq_file->do_lu);
 310                 d->opd_last_used_seq_file = NULL;
 311         }
 312 }
 313
 314 static int osp_disconnect(struct osp_device *d)
 315 {
 316         struct obd_import *imp;
 317         int rc = 0;
 318
 319         imp = d->opd_obd->u.cli.cl_import;
 320
 321         /* Mark import deactivated now, so we don't try to reconnect if any
 322          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 323          * fully deactivate the import, or that would drop all requests. */
 324         LASSERT(imp != NULL);
 325         spin_lock(&imp->imp_lock);
 326         imp->imp_deactive = 1;
 327         spin_unlock(&imp->imp_lock);
 328
 329         ptlrpc_deactivate_import(imp);
 330
 331         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 332          * delete it regardless.  (It's safe to delete an import that was
 333          * never added.) */
 334         (void)ptlrpc_pinger_del_import(imp);
 335
 336         rc = ptlrpc_disconnect_import(imp, 0);
 337         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
 338                 rc = 0;
 339         if (rc)
 340                 CERROR("%s: can't disconnect: rc = %d\n",
 341                        d->opd_obd->obd_name, rc);
 342
 343         ptlrpc_invalidate_import(imp);
 344
 345         RETURN(rc);
 346 }
 347
 348 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
 349 {
 350         int                      rc = 0;
 351         ENTRY;
 352
 353         LASSERT(env);
 354         /* release last_used file */
 355         if (!d->opd_connect_mdt)
 356                 osp_last_used_fini(env, d);
 357
 358         rc = osp_disconnect(d);
 359
 360         if (!d->opd_connect_mdt) {
 361                 /* stop precreate thread */
 362                 osp_precreate_fini(d);
 363
 364                 /* stop sync thread */
 365                 osp_sync_fini(d);
 366         }
 367
 368         obd_fid_fini(d->opd_obd);
 369
 370         RETURN(rc);
 371 }
 372
 373 static int osp_process_config(const struct lu_env *env,
 374                               struct lu_device *dev, struct lustre_cfg *lcfg)
 375 {
 376         struct osp_device               *d = lu2osp_dev(dev);
 377         struct lprocfs_static_vars       lvars = { 0 };
 378         int                              rc;
 379
 380         ENTRY;
 381
 382         switch (lcfg->lcfg_command) {
 383         case LCFG_CLEANUP:
 384                 lu_dev_del_linkage(dev->ld_site, dev);
 385                 rc = osp_shutdown(env, d);
 386                 break;
 387         case LCFG_PARAM:
 388                 lprocfs_osp_init_vars(&lvars);
 389
 390                 LASSERT(d->opd_obd);
 391                 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
 392                                               lcfg, d->opd_obd);
 393                 if (rc > 0)
 394                         rc = 0;
 395                 if (rc == -ENOSYS) {
 396                         /* class_process_proc_param() haven't found matching
 397                          * parameter and returned ENOSYS so that layer(s)
 398                          * below could use that. But OSP is the bottom, so
 399                          * just ignore it */
 400                         CERROR("%s: unknown param %s\n",
 401                                (char *)lustre_cfg_string(lcfg, 0),
 402                                (char *)lustre_cfg_string(lcfg, 1));
 403                         rc = 0;
 404                 }
 405                 break;
 406         default:
 407                 CERROR("%s: unknown command %u\n",
 408                        (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
 409                 rc = 0;
 410                 break;
 411         }
 412
 413         RETURN(rc);
 414 }
 415
 416 static int osp_recovery_complete(const struct lu_env *env,
 417                                  struct lu_device *dev)
 418 {
 419         struct osp_device       *osp = lu2osp_dev(dev);
 420         int                      rc = 0;
 421
 422         ENTRY;
 423         osp->opd_recovery_completed = 1;
 424         if (!osp->opd_connect_mdt)
 425                 cfs_waitq_signal(&osp->opd_pre_waitq);
 426         RETURN(rc);
 427 }
 428
 429 const struct lu_device_operations osp_lu_ops = {
 430         .ldo_object_alloc       = osp_object_alloc,
 431         .ldo_process_config     = osp_process_config,
 432         .ldo_recovery_complete  = osp_recovery_complete,
 433 };
 434
 435 /**
 436  * provides with statfs from corresponded OST
 437  *
 438  */
 439 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
 440                       struct obd_statfs *sfs)
 441 {
 442         struct osp_device *d = dt2osp_dev(dev);
 443
 444         ENTRY;
 445
 446         if (unlikely(d->opd_imp_active == 0))
 447                 RETURN(-ENOTCONN);
 448
 449         /* return recently updated data */
 450         *sfs = d->opd_statfs;
 451
 452         /*
 453          * layer above osp (usually lod) can use ffree to estimate
 454          * how many objects are available for immediate creation
 455          */
 456
 457         spin_lock(&d->opd_pre_lock);
 458         LASSERTF(fid_seq(&d->opd_pre_last_created_fid) ==
 459                  fid_seq(&d->opd_pre_used_fid),
 460                  "last_created "DFID", next_fid "DFID"\n",
 461                  PFID(&d->opd_pre_last_created_fid),
 462                  PFID(&d->opd_pre_used_fid));
 463         sfs->os_fprecreated = fid_oid(&d->opd_pre_last_created_fid) -
 464                               fid_oid(&d->opd_pre_used_fid);
 465         sfs->os_fprecreated -= d->opd_pre_reserved;
 466         spin_unlock(&d->opd_pre_lock);
 467
 468         LASSERT(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2);
 469
 470         CDEBUG(D_OTHER, "%s: "LPU64" blocks, "LPU64" free, "LPU64" avail, "
 471                LPU64" files, "LPU64" free files\n", d->opd_obd->obd_name,
 472                sfs->os_blocks, sfs->os_bfree, sfs->os_bavail,
 473                sfs->os_files, sfs->os_ffree);
 474         RETURN(0);
 475 }
 476
 477 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
 478 {
 479         ENTRY;
 480
 481         /*
 482          * XXX: wake up sync thread, command it to start flushing asap?
 483          */
 484
 485         RETURN(0);
 486 }
 487
 488 const struct dt_device_operations osp_dt_ops = {
 489         .dt_statfs      = osp_statfs,
 490         .dt_sync        = osp_sync,
 491 };
 492
 493 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *m,
 494                               const char *nextdev)
 495 {
 496         struct obd_connect_data *data = NULL;
 497         struct obd_device       *obd;
 498         int                      rc;
 499
 500         ENTRY;
 501
 502         LASSERT(m->opd_storage_exp == NULL);
 503
 504         OBD_ALLOC_PTR(data);
 505         if (data == NULL)
 506                 RETURN(-ENOMEM);
 507
 508         obd = class_name2obd(nextdev);
 509         if (obd == NULL) {
 510                 CERROR("%s: can't locate next device: %s\n",
 511                        m->opd_obd->obd_name, nextdev);
 512                 GOTO(out, rc = -ENOTCONN);
 513         }
 514
 515         rc = obd_connect(env, &m->opd_storage_exp, obd, &obd->obd_uuid, data,
 516                          NULL);
 517         if (rc) {
 518                 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
 519                        m->opd_obd->obd_name, nextdev, rc);
 520                 GOTO(out, rc);
 521         }
 522
 523         m->opd_dt_dev.dd_lu_dev.ld_site =
 524                 m->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
 525         LASSERT(m->opd_dt_dev.dd_lu_dev.ld_site);
 526         m->opd_storage = lu2dt_dev(m->opd_storage_exp->exp_obd->obd_lu_dev);
 527
 528 out:
 529         OBD_FREE_PTR(data);
 530         RETURN(rc);
 531 }
 532
 533 static int osp_init0(const struct lu_env *env, struct osp_device *m,
 534                      struct lu_device_type *ldt, struct lustre_cfg *cfg)
 535 {
 536         struct obd_device       *obd;
 537         struct obd_import       *imp;
 538         class_uuid_t            uuid;
 539         char                    *src, *tgt, *mdt, *osdname = NULL;
 540         int                     rc, idx;
 541
 542         ENTRY;
 543
 544         obd = class_name2obd(lustre_cfg_string(cfg, 0));
 545         if (obd == NULL) {
 546                 CERROR("Cannot find obd with name %s\n",
 547                        lustre_cfg_string(cfg, 0));
 548                 RETURN(-ENODEV);
 549         }
 550         m->opd_obd = obd;
 551
 552         /* There is no record in the MDT configuration for the local disk
 553          * device, so we have to extract this from elsewhere in the profile.
 554          * The only information we get at setup is from the OSC records:
 555          * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 556          * Note that 1.8 generated configs are missing the -MDTxxxx part.
 557          * We need to reconstruct the name of the underlying OSD from this:
 558          * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".  We
 559          * also need to determine the OST index from this - will be used
 560          * to calculate the offset in shared lov_objids file later */
 561
 562         src = lustre_cfg_string(cfg, 0);
 563         if (src == NULL)
 564                 RETURN(-EINVAL);
 565
 566         tgt = strrchr(src, '-');
 567         if (tgt == NULL) {
 568                 CERROR("%s: invalid target name %s\n",
 569                        m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 570                 RETURN(-EINVAL);
 571         }
 572
 573         if (strncmp(tgt, "-osc", 4) == 0) {
 574                 /* Old OSC name fsname-OSTXXXX-osc */
 575                 for (tgt--; tgt > src && *tgt != '-'; tgt--)
 576                         ;
 577                 if (tgt == src) {
 578                         CERROR("%s: invalid target name %s\n",
 579                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 580                         RETURN(-EINVAL);
 581                 }
 582
 583                 if (strncmp(tgt, "-OST", 4) != 0) {
 584                         CERROR("%s: invalid target name %s\n",
 585                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 586                         RETURN(-EINVAL);
 587                 }
 588
 589                 idx = simple_strtol(tgt + 4, &mdt, 16);
 590                 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
 591                         CERROR("%s: invalid OST index in '%s'\n",
 592                                m->opd_obd->obd_name, src);
 593                         RETURN(-EINVAL);
 594                 }
 595                 m->opd_index = idx;
 596                 idx = tgt - src;
 597         } else {
 598                 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
 599                 if (strncmp(tgt, "-MDT", 4) != 0 &&
 600                          strncmp(tgt, "-OST", 4) != 0) {
 601                         CERROR("%s: invalid target name %s\n",
 602                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 603                         RETURN(-EINVAL);
 604                 }
 605
 606                 if (tgt - src <= 12) {
 607                         CERROR("%s: invalid target name %s\n",
 608                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 609                         RETURN(-EINVAL);
 610                 }
 611
 612                 if (strncmp(tgt - 12, "-MDT", 4) == 0)
 613                         m->opd_connect_mdt = 1;
 614
 615                 idx = simple_strtol(tgt - 8, &mdt, 16);
 616                 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
 617                         CERROR("%s: invalid OST index in '%s'\n",
 618                                m->opd_obd->obd_name, src);
 619                         RETURN(-EINVAL);
 620                 }
 621
 622                 m->opd_index = idx;
 623                 idx = tgt - src - 12;
 624         }
 625         /* check the fsname length, and after this everything else will fit */
 626         if (idx > MTI_NAME_MAXLEN) {
 627                 CERROR("%s: fsname too long in '%s'\n",
 628                        m->opd_obd->obd_name, src);
 629                 RETURN(-EINVAL);
 630         }
 631
 632         OBD_ALLOC(osdname, MAX_OBD_NAME);
 633         if (osdname == NULL)
 634                 RETURN(-ENOMEM);
 635
 636         memcpy(osdname, src, idx); /* copy just the fsname part */
 637         osdname[idx] = '\0';
 638
 639         mdt = strstr(mdt, "-MDT");
 640         if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
 641                 strcat(osdname, "-MDT0000");
 642         else
 643                 strcat(osdname, mdt);
 644         strcat(osdname, "-osd");
 645         CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
 646
 647         if (m->opd_connect_mdt) {
 648                 struct client_obd *cli = &m->opd_obd->u.cli;
 649
 650                 OBD_ALLOC(cli->cl_rpc_lock, sizeof(*cli->cl_rpc_lock));
 651                 if (!cli->cl_rpc_lock)
 652                         RETURN(-ENOMEM);
 653                 osp_init_rpc_lock(cli->cl_rpc_lock);
 654         }
 655
 656         m->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
 657         m->opd_dt_dev.dd_ops = &osp_dt_ops;
 658         obd->obd_lu_dev = &m->opd_dt_dev.dd_lu_dev;
 659
 660         rc = osp_connect_to_osd(env, m, osdname);
 661         if (rc)
 662                 GOTO(out_fini, rc);
 663
 664         rc = ptlrpcd_addref();
 665         if (rc)
 666                 GOTO(out_disconnect, rc);
 667
 668         rc = client_obd_setup(obd, cfg);
 669         if (rc) {
 670                 CERROR("%s: can't setup obd: %d\n", m->opd_obd->obd_name, rc);
 671                 GOTO(out_ref, rc);
 672         }
 673
 674         osp_lprocfs_init(m);
 675
 676         if (!m->opd_connect_mdt) {
 677                 /* Initialize last id from the storage - will be
 678                  * used in orphan cleanup. */
 679                 rc = osp_last_used_init(env, m);
 680                 if (rc)
 681                         GOTO(out_proc, rc);
 682                 /* Initialize precreation thread, it handles new
 683                  * connections as well. */
 684                 rc = osp_init_precreate(m);
 685                 if (rc)
 686                         GOTO(out_last_used, rc);
 687                 /*
 688                  * Initialize synhronization mechanism taking
 689                  * care of propogating changes to OST in near
 690                  * transactional manner.
 691                  */
 692                 rc = osp_sync_init(env, m);
 693                 if (rc)
 694                         GOTO(out_precreat, rc);
 695
 696                 rc = obd_fid_init(m->opd_obd, NULL, LUSTRE_SEQ_DATA);
 697                 if (rc) {
 698                         CERROR("%s: fid init error: rc = %d\n",
 699                                m->opd_obd->obd_name, rc);
 700                         GOTO(out, rc);
 701                 }
 702         }
 703         /*
 704          * Initiate connect to OST
 705          */
 706         ll_generate_random_uuid(uuid);
 707         class_uuid_unparse(uuid, &m->opd_cluuid);
 708
 709         imp = obd->u.cli.cl_import;
 710
 711         rc = ptlrpc_init_import(imp);
 712         if (rc)
 713                 GOTO(out, rc);
 714         if (osdname)
 715                 OBD_FREE(osdname, MAX_OBD_NAME);
 716         RETURN(0);
 717
 718 out:
 719         if (!m->opd_connect_mdt)
 720                 /* stop sync thread */
 721                 osp_sync_fini(m);
 722 out_precreat:
 723         /* stop precreate thread */
 724         if (!m->opd_connect_mdt)
 725                 osp_precreate_fini(m);
 726 out_last_used:
 727         osp_last_used_fini(env, m);
 728 out_proc:
 729         ptlrpc_lprocfs_unregister_obd(obd);
 730         lprocfs_obd_cleanup(obd);
 731         class_destroy_import(obd->u.cli.cl_import);
 732         client_obd_cleanup(obd);
 733 out_ref:
 734         ptlrpcd_decref();
 735 out_disconnect:
 736         if (m->opd_connect_mdt) {
 737                 struct client_obd *cli = &m->opd_obd->u.cli;
 738                 if (cli->cl_rpc_lock != NULL) {
 739                         OBD_FREE_PTR(cli->cl_rpc_lock);
 740                         cli->cl_rpc_lock = NULL;
 741                 }
 742         }
 743         obd_disconnect(m->opd_storage_exp);
 744 out_fini:
 745         if (osdname)
 746                 OBD_FREE(osdname, MAX_OBD_NAME);
 747         RETURN(rc);
 748 }
 749
 750 static struct lu_device *osp_device_free(const struct lu_env *env,
 751                                          struct lu_device *lu)
 752 {
 753         struct osp_device *m = lu2osp_dev(lu);
 754
 755         ENTRY;
 756
 757         if (cfs_atomic_read(&lu->ld_ref) && lu->ld_site) {
 758                 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
 759                 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
 760         }
 761         dt_device_fini(&m->opd_dt_dev);
 762         OBD_FREE_PTR(m);
 763         RETURN(NULL);
 764 }
 765
 766 static struct lu_device *osp_device_alloc(const struct lu_env *env,
 767                                           struct lu_device_type *t,
 768                                           struct lustre_cfg *lcfg)
 769 {
 770         struct osp_device *m;
 771         struct lu_device  *l;
 772
 773         OBD_ALLOC_PTR(m);
 774         if (m == NULL) {
 775                 l = ERR_PTR(-ENOMEM);
 776         } else {
 777                 int rc;
 778
 779                 l = osp2lu_dev(m);
 780                 dt_device_init(&m->opd_dt_dev, t);
 781                 rc = osp_init0(env, m, t, lcfg);
 782                 if (rc != 0) {
 783                         osp_device_free(env, l);
 784                         l = ERR_PTR(rc);
 785                 }
 786         }
 787         return l;
 788 }
 789
 790 static struct lu_device *osp_device_fini(const struct lu_env *env,
 791                                          struct lu_device *d)
 792 {
 793         struct osp_device *m = lu2osp_dev(d);
 794         struct obd_import *imp;
 795         int                rc;
 796
 797         ENTRY;
 798
 799         if (m->opd_storage_exp)
 800                 obd_disconnect(m->opd_storage_exp);
 801
 802         imp = m->opd_obd->u.cli.cl_import;
 803
 804         if (imp->imp_rq_pool) {
 805                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 806                 imp->imp_rq_pool = NULL;
 807         }
 808
 809         obd_cleanup_client_import(m->opd_obd);
 810
 811         if (m->opd_symlink)
 812                 lprocfs_remove(&m->opd_symlink);
 813
 814         LASSERT(m->opd_obd);
 815         ptlrpc_lprocfs_unregister_obd(m->opd_obd);
 816         lprocfs_obd_cleanup(m->opd_obd);
 817
 818         if (m->opd_connect_mdt) {
 819                 struct client_obd *cli = &m->opd_obd->u.cli;
 820                 if (cli->cl_rpc_lock != NULL) {
 821                         OBD_FREE_PTR(cli->cl_rpc_lock);
 822                         cli->cl_rpc_lock = NULL;
 823                 }
 824         }
 825
 826         rc = client_obd_cleanup(m->opd_obd);
 827         LASSERTF(rc == 0, "error %d\n", rc);
 828
 829         ptlrpcd_decref();
 830
 831         RETURN(NULL);
 832 }
 833
 834 static int osp_reconnect(const struct lu_env *env,
 835                          struct obd_export *exp, struct obd_device *obd,
 836                          struct obd_uuid *cluuid,
 837                          struct obd_connect_data *data,
 838                          void *localdata)
 839 {
 840         return 0;
 841 }
 842
 843 /*
 844  * we use exports to track all LOD users
 845  */
 846 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
 847                            struct obd_device *obd, struct obd_uuid *cluuid,
 848                            struct obd_connect_data *data, void *localdata)
 849 {
 850         struct osp_device       *osp = lu2osp_dev(obd->obd_lu_dev);
 851         struct obd_connect_data *ocd;
 852         struct obd_import       *imp;
 853         struct lustre_handle     conn;
 854         int                      rc;
 855
 856         ENTRY;
 857
 858         CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
 859
 860         rc = class_connect(&conn, obd, cluuid);
 861         if (rc)
 862                 RETURN(rc);
 863
 864         *exp = class_conn2export(&conn);
 865         /* Why should there ever be more than 1 connect? */
 866         osp->opd_connects++;
 867         LASSERT(osp->opd_connects == 1);
 868
 869         osp->opd_exp = *exp;
 870
 871         imp = osp->opd_obd->u.cli.cl_import;
 872         imp->imp_dlm_handle = conn;
 873
 874         LASSERT(data != NULL);
 875         LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
 876         ocd = &imp->imp_connect_data;
 877         *ocd = *data;
 878
 879         imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
 880
 881         ocd->ocd_version = LUSTRE_VERSION_CODE;
 882         ocd->ocd_index = data->ocd_index;
 883         imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
 884
 885         rc = ptlrpc_connect_import(imp);
 886         if (rc) {
 887                 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
 888                 GOTO(out, rc);
 889         }
 890
 891         ptlrpc_pinger_add_import(imp);
 892
 893         if (osp->opd_connect_mdt && data->ocd_index == 0) {
 894                 /* set seq controller export for MDC0 if exists */
 895                 struct seq_server_site *ss;
 896
 897                 ss = lu_site2seq(osp2lu_dev(osp)->ld_site);
 898                 ss->ss_control_exp = class_export_get(*exp);
 899                 ss->ss_server_fld->lsf_control_exp = *exp;
 900         }
 901 out:
 902         RETURN(rc);
 903 }
 904
 905 /*
 906  * once last export (we don't count self-export) disappeared
 907  * osp can be released
 908  */
 909 static int osp_obd_disconnect(struct obd_export *exp)
 910 {
 911         struct obd_device *obd = exp->exp_obd;
 912         struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
 913         int                rc;
 914         ENTRY;
 915
 916         /* Only disconnect the underlying layers on the final disconnect. */
 917         LASSERT(osp->opd_connects == 1);
 918         osp->opd_connects--;
 919
 920         rc = class_disconnect(exp);
 921         if (rc) {
 922                 CERROR("%s: class disconnect error: rc = %d\n",
 923                        obd->obd_name, rc);
 924                 RETURN(rc);
 925         }
 926
 927         /* destroy the device */
 928         class_manual_cleanup(obd);
 929
 930         RETURN(rc);
 931 }
 932
 933 /*
 934  * lprocfs helpers still use OBD API, let's keep obd_statfs() support
 935  */
 936 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
 937                           struct obd_statfs *osfs, __u64 max_age, __u32 flags)
 938 {
 939         struct obd_statfs       *msfs;
 940         struct ptlrpc_request   *req;
 941         struct obd_import       *imp = NULL;
 942         int                      rc;
 943
 944         ENTRY;
 945
 946         /* Since the request might also come from lprocfs, so we need
 947          * sync this with client_disconnect_export Bug15684 */
 948         down_read(&exp->exp_obd->u.cli.cl_sem);
 949         if (exp->exp_obd->u.cli.cl_import)
 950                 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
 951         up_read(&exp->exp_obd->u.cli.cl_sem);
 952         if (!imp)
 953                 RETURN(-ENODEV);
 954
 955         /* We could possibly pass max_age in the request (as an absolute
 956          * timestamp or a "seconds.usec ago") so the target can avoid doing
 957          * extra calls into the filesystem if that isn't necessary (e.g.
 958          * during mount that would help a bit).  Having relative timestamps
 959          * is not so great if request processing is slow, while absolute
 960          * timestamps are not ideal because they need time synchronization. */
 961         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
 962
 963         class_import_put(imp);
 964
 965         if (req == NULL)
 966                 RETURN(-ENOMEM);
 967
 968         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
 969         if (rc) {
 970                 ptlrpc_request_free(req);
 971                 RETURN(rc);
 972         }
 973         ptlrpc_request_set_replen(req);
 974         req->rq_request_portal = OST_CREATE_PORTAL;
 975         ptlrpc_at_set_req_timeout(req);
 976
 977         if (flags & OBD_STATFS_NODELAY) {
 978                 /* procfs requests not want stat in wait for avoid deadlock */
 979                 req->rq_no_resend = 1;
 980                 req->rq_no_delay = 1;
 981         }
 982
 983         rc = ptlrpc_queue_wait(req);
 984         if (rc)
 985                 GOTO(out, rc);
 986
 987         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
 988         if (msfs == NULL)
 989                 GOTO(out, rc = -EPROTO);
 990
 991         *osfs = *msfs;
 992
 993         EXIT;
 994 out:
 995         ptlrpc_req_finished(req);
 996         return rc;
 997 }
 998
 999 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
1000                             enum obd_import_event event)
1001 {
1002         struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1003
1004         switch (event) {
1005         case IMP_EVENT_DISCON:
1006                 d->opd_got_disconnected = 1;
1007                 d->opd_imp_connected = 0;
1008                 if (d->opd_connect_mdt)
1009                         break;
1010                 osp_pre_update_status(d, -ENODEV);
1011                 cfs_waitq_signal(&d->opd_pre_waitq);
1012                 CDEBUG(D_HA, "got disconnected\n");
1013                 break;
1014         case IMP_EVENT_INACTIVE:
1015                 d->opd_imp_active = 0;
1016                 if (d->opd_connect_mdt)
1017                         break;
1018                 osp_pre_update_status(d, -ENODEV);
1019                 cfs_waitq_signal(&d->opd_pre_waitq);
1020                 CDEBUG(D_HA, "got inactive\n");
1021                 break;
1022         case IMP_EVENT_ACTIVE:
1023                 d->opd_imp_active = 1;
1024                 if (d->opd_got_disconnected)
1025                         d->opd_new_connection = 1;
1026                 d->opd_imp_connected = 1;
1027                 d->opd_imp_seen_connected = 1;
1028                 if (d->opd_connect_mdt)
1029                         break;
1030                 cfs_waitq_signal(&d->opd_pre_waitq);
1031                 __osp_sync_check_for_work(d);
1032                 CDEBUG(D_HA, "got connected\n");
1033                 break;
1034         case IMP_EVENT_INVALIDATE:
1035                 if (obd->obd_namespace == NULL)
1036                         break;
1037                 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1038                 break;
1039         case IMP_EVENT_OCD:
1040         case IMP_EVENT_DEACTIVATE:
1041         case IMP_EVENT_ACTIVATE:
1042                 break;
1043         default:
1044                 CERROR("%s: unsupported import event: %#x\n",
1045                        obd->obd_name, event);
1046         }
1047         return 0;
1048 }
1049
1050 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1051                          void *karg, void *uarg)
1052 {
1053         struct obd_device       *obd = exp->exp_obd;
1054         struct osp_device       *d;
1055         struct obd_ioctl_data   *data = karg;
1056         int                      rc = 0;
1057
1058         ENTRY;
1059
1060         LASSERT(obd->obd_lu_dev);
1061         d = lu2osp_dev(obd->obd_lu_dev);
1062         LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1063
1064         if (!cfs_try_module_get(THIS_MODULE)) {
1065                 CERROR("%s: can't get module. Is it alive?", obd->obd_name);
1066                 return -EINVAL;
1067         }
1068
1069         switch (cmd) {
1070         case OBD_IOC_CLIENT_RECOVER:
1071                 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1072                                            data->ioc_inlbuf1, 0);
1073                 if (rc > 0)
1074                         rc = 0;
1075                 break;
1076         case IOC_OSC_SET_ACTIVE:
1077                 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1078                                               data->ioc_offset);
1079                 break;
1080         case OBD_IOC_PING_TARGET:
1081                 rc = ptlrpc_obd_ping(obd);
1082                 break;
1083         default:
1084                 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1085                        cmd, cfs_curproc_comm());
1086                 rc = -ENOTTY;
1087         }
1088         cfs_module_put(THIS_MODULE);
1089         return rc;
1090 }
1091
1092 static int osp_obd_health_check(const struct lu_env *env,
1093                                 struct obd_device *obd)
1094 {
1095         struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1096
1097         ENTRY;
1098
1099         /*
1100          * 1.8/2.0 behaviour is that OST being connected once at least
1101          * is considired "healthy". and one "healty" OST is enough to
1102          * allow lustre clients to connect to MDS
1103          */
1104         LASSERT(d);
1105         RETURN(!d->opd_imp_seen_connected);
1106 }
1107
1108 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1109 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1110 static void osp_key_exit(const struct lu_context *ctx,
1111                          struct lu_context_key *key, void *data)
1112 {
1113         struct osp_thread_info *info = data;
1114
1115         info->osi_attr.la_valid = 0;
1116 }
1117
1118 struct lu_context_key osp_thread_key = {
1119         .lct_tags = LCT_MD_THREAD,
1120         .lct_init = osp_key_init,
1121         .lct_fini = osp_key_fini,
1122         .lct_exit = osp_key_exit
1123 };
1124
1125 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1126 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1127
1128 struct lu_context_key osp_txn_key = {
1129         .lct_tags = LCT_OSP_THREAD,
1130         .lct_init = osp_txn_key_init,
1131         .lct_fini = osp_txn_key_fini
1132 };
1133 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1134
1135 static struct lu_device_type_operations osp_device_type_ops = {
1136         .ldto_init           = osp_type_init,
1137         .ldto_fini           = osp_type_fini,
1138
1139         .ldto_start          = osp_type_start,
1140         .ldto_stop           = osp_type_stop,
1141
1142         .ldto_device_alloc   = osp_device_alloc,
1143         .ldto_device_free    = osp_device_free,
1144
1145         .ldto_device_fini    = osp_device_fini
1146 };
1147
1148 static struct lu_device_type osp_device_type = {
1149         .ldt_tags     = LU_DEVICE_DT,
1150         .ldt_name     = LUSTRE_OSP_NAME,
1151         .ldt_ops      = &osp_device_type_ops,
1152         .ldt_ctx_tags = LCT_MD_THREAD
1153 };
1154
1155 static struct obd_ops osp_obd_device_ops = {
1156         .o_owner        = THIS_MODULE,
1157         .o_add_conn     = client_import_add_conn,
1158         .o_del_conn     = client_import_del_conn,
1159         .o_reconnect    = osp_reconnect,
1160         .o_connect      = osp_obd_connect,
1161         .o_disconnect   = osp_obd_disconnect,
1162         .o_health_check = osp_obd_health_check,
1163         .o_import_event = osp_import_event,
1164         .o_iocontrol    = osp_iocontrol,
1165         .o_statfs       = osp_obd_statfs,
1166         .o_fid_init     = client_fid_init,
1167         .o_fid_fini     = client_fid_fini,
1168 };
1169
1170 struct llog_operations osp_mds_ost_orig_logops;
1171
1172 static int __init osp_mod_init(void)
1173 {
1174         struct lprocfs_static_vars       lvars;
1175         cfs_proc_dir_entry_t            *osc_proc_dir;
1176         int                              rc;
1177
1178         rc = lu_kmem_init(osp_caches);
1179         if (rc)
1180                 return rc;
1181
1182         lprocfs_osp_init_vars(&lvars);
1183
1184         rc = class_register_type(&osp_obd_device_ops, NULL, lvars.module_vars,
1185                                  LUSTRE_OSP_NAME, &osp_device_type);
1186
1187         /* create "osc" entry in procfs for compatibility purposes */
1188         if (rc != 0) {
1189                 lu_kmem_fini(osp_caches);
1190                 return rc;
1191         }
1192
1193         lprocfs_lwp_init_vars(&lvars);
1194
1195         rc = class_register_type(&lwp_obd_device_ops, NULL, lvars.module_vars,
1196                                  LUSTRE_LWP_NAME, &lwp_device_type);
1197         if (rc != 0) {
1198                 class_unregister_type(LUSTRE_OSP_NAME);
1199                 lu_kmem_fini(osp_caches);
1200                 return rc;
1201         }
1202
1203         /* Note: add_rec/delcare_add_rec will be only used by catalogs */
1204         osp_mds_ost_orig_logops = llog_osd_ops;
1205         osp_mds_ost_orig_logops.lop_add = llog_cat_add_rec;
1206         osp_mds_ost_orig_logops.lop_declare_add = llog_cat_declare_add_rec;
1207
1208         osc_proc_dir = lprocfs_srch(proc_lustre_root, "osc");
1209         if (osc_proc_dir == NULL) {
1210                 osc_proc_dir = lprocfs_register("osc", proc_lustre_root, NULL,
1211                                                 NULL);
1212                 if (IS_ERR(osc_proc_dir))
1213                         CERROR("osp: can't create compat entry \"osc\": %d\n",
1214                                (int) PTR_ERR(osc_proc_dir));
1215         }
1216         return rc;
1217 }
1218
1219 static void __exit osp_mod_exit(void)
1220 {
1221         lprocfs_try_remove_proc_entry("osc", proc_lustre_root);
1222
1223         class_unregister_type(LUSTRE_LWP_NAME);
1224         class_unregister_type(LUSTRE_OSP_NAME);
1225         lu_kmem_fini(osp_caches);
1226 }
1227
1228 MODULE_AUTHOR("Intel, Inc. <http://www.intel.com/>");
1229 MODULE_DESCRIPTION("Lustre OST Proxy Device ("LUSTRE_OSP_NAME")");
1230 MODULE_LICENSE("GPL");
1231
1232 cfs_module(osp, LUSTRE_VERSION_STRING, osp_mod_init, osp_mod_exit);