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         .dt_trans_start = osp_trans_start,
 492         .dt_trans_stop  = osp_trans_stop,
 493 };
 494
 495 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *m,
 496                               const char *nextdev)
 497 {
 498         struct obd_connect_data *data = NULL;
 499         struct obd_device       *obd;
 500         int                      rc;
 501
 502         ENTRY;
 503
 504         LASSERT(m->opd_storage_exp == NULL);
 505
 506         OBD_ALLOC_PTR(data);
 507         if (data == NULL)
 508                 RETURN(-ENOMEM);
 509
 510         obd = class_name2obd(nextdev);
 511         if (obd == NULL) {
 512                 CERROR("%s: can't locate next device: %s\n",
 513                        m->opd_obd->obd_name, nextdev);
 514                 GOTO(out, rc = -ENOTCONN);
 515         }
 516
 517         rc = obd_connect(env, &m->opd_storage_exp, obd, &obd->obd_uuid, data,
 518                          NULL);
 519         if (rc) {
 520                 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
 521                        m->opd_obd->obd_name, nextdev, rc);
 522                 GOTO(out, rc);
 523         }
 524
 525         m->opd_dt_dev.dd_lu_dev.ld_site =
 526                 m->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
 527         LASSERT(m->opd_dt_dev.dd_lu_dev.ld_site);
 528         m->opd_storage = lu2dt_dev(m->opd_storage_exp->exp_obd->obd_lu_dev);
 529
 530 out:
 531         OBD_FREE_PTR(data);
 532         RETURN(rc);
 533 }
 534
 535 static int osp_init0(const struct lu_env *env, struct osp_device *m,
 536                      struct lu_device_type *ldt, struct lustre_cfg *cfg)
 537 {
 538         struct obd_device       *obd;
 539         struct obd_import       *imp;
 540         class_uuid_t            uuid;
 541         char                    *src, *tgt, *mdt, *osdname = NULL;
 542         int                     rc, idx;
 543
 544         ENTRY;
 545
 546         obd = class_name2obd(lustre_cfg_string(cfg, 0));
 547         if (obd == NULL) {
 548                 CERROR("Cannot find obd with name %s\n",
 549                        lustre_cfg_string(cfg, 0));
 550                 RETURN(-ENODEV);
 551         }
 552         m->opd_obd = obd;
 553
 554         /* There is no record in the MDT configuration for the local disk
 555          * device, so we have to extract this from elsewhere in the profile.
 556          * The only information we get at setup is from the OSC records:
 557          * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 558          * Note that 1.8 generated configs are missing the -MDTxxxx part.
 559          * We need to reconstruct the name of the underlying OSD from this:
 560          * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".  We
 561          * also need to determine the OST index from this - will be used
 562          * to calculate the offset in shared lov_objids file later */
 563
 564         src = lustre_cfg_string(cfg, 0);
 565         if (src == NULL)
 566                 RETURN(-EINVAL);
 567
 568         tgt = strrchr(src, '-');
 569         if (tgt == NULL) {
 570                 CERROR("%s: invalid target name %s\n",
 571                        m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 572                 RETURN(-EINVAL);
 573         }
 574
 575         if (strncmp(tgt, "-osc", 4) == 0) {
 576                 /* Old OSC name fsname-OSTXXXX-osc */
 577                 for (tgt--; tgt > src && *tgt != '-'; tgt--)
 578                         ;
 579                 if (tgt == src) {
 580                         CERROR("%s: invalid target name %s\n",
 581                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 582                         RETURN(-EINVAL);
 583                 }
 584
 585                 if (strncmp(tgt, "-OST", 4) != 0) {
 586                         CERROR("%s: invalid target name %s\n",
 587                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 588                         RETURN(-EINVAL);
 589                 }
 590
 591                 idx = simple_strtol(tgt + 4, &mdt, 16);
 592                 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
 593                         CERROR("%s: invalid OST index in '%s'\n",
 594                                m->opd_obd->obd_name, src);
 595                         RETURN(-EINVAL);
 596                 }
 597                 m->opd_index = idx;
 598                 idx = tgt - src;
 599         } else {
 600                 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
 601                 if (strncmp(tgt, "-MDT", 4) != 0 &&
 602                          strncmp(tgt, "-OST", 4) != 0) {
 603                         CERROR("%s: invalid target name %s\n",
 604                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 605                         RETURN(-EINVAL);
 606                 }
 607
 608                 if (tgt - src <= 12) {
 609                         CERROR("%s: invalid target name %s\n",
 610                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
 611                         RETURN(-EINVAL);
 612                 }
 613
 614                 if (strncmp(tgt - 12, "-MDT", 4) == 0)
 615                         m->opd_connect_mdt = 1;
 616
 617                 idx = simple_strtol(tgt - 8, &mdt, 16);
 618                 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
 619                         CERROR("%s: invalid OST index in '%s'\n",
 620                                m->opd_obd->obd_name, src);
 621                         RETURN(-EINVAL);
 622                 }
 623
 624                 m->opd_index = idx;
 625                 idx = tgt - src - 12;
 626         }
 627         /* check the fsname length, and after this everything else will fit */
 628         if (idx > MTI_NAME_MAXLEN) {
 629                 CERROR("%s: fsname too long in '%s'\n",
 630                        m->opd_obd->obd_name, src);
 631                 RETURN(-EINVAL);
 632         }
 633
 634         OBD_ALLOC(osdname, MAX_OBD_NAME);
 635         if (osdname == NULL)
 636                 RETURN(-ENOMEM);
 637
 638         memcpy(osdname, src, idx); /* copy just the fsname part */
 639         osdname[idx] = '\0';
 640
 641         mdt = strstr(mdt, "-MDT");
 642         if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
 643                 strcat(osdname, "-MDT0000");
 644         else
 645                 strcat(osdname, mdt);
 646         strcat(osdname, "-osd");
 647         CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
 648
 649         if (m->opd_connect_mdt) {
 650                 struct client_obd *cli = &m->opd_obd->u.cli;
 651
 652                 OBD_ALLOC(cli->cl_rpc_lock, sizeof(*cli->cl_rpc_lock));
 653                 if (!cli->cl_rpc_lock)
 654                         RETURN(-ENOMEM);
 655                 osp_init_rpc_lock(cli->cl_rpc_lock);
 656         }
 657
 658         m->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
 659         m->opd_dt_dev.dd_ops = &osp_dt_ops;
 660         obd->obd_lu_dev = &m->opd_dt_dev.dd_lu_dev;
 661
 662         rc = osp_connect_to_osd(env, m, osdname);
 663         if (rc)
 664                 GOTO(out_fini, rc);
 665
 666         rc = ptlrpcd_addref();
 667         if (rc)
 668                 GOTO(out_disconnect, rc);
 669
 670         rc = client_obd_setup(obd, cfg);
 671         if (rc) {
 672                 CERROR("%s: can't setup obd: %d\n", m->opd_obd->obd_name, rc);
 673                 GOTO(out_ref, rc);
 674         }
 675
 676         osp_lprocfs_init(m);
 677
 678         if (!m->opd_connect_mdt) {
 679                 /* Initialize last id from the storage - will be
 680                  * used in orphan cleanup. */
 681                 rc = osp_last_used_init(env, m);
 682                 if (rc)
 683                         GOTO(out_proc, rc);
 684                 /* Initialize precreation thread, it handles new
 685                  * connections as well. */
 686                 rc = osp_init_precreate(m);
 687                 if (rc)
 688                         GOTO(out_last_used, rc);
 689                 /*
 690                  * Initialize synhronization mechanism taking
 691                  * care of propogating changes to OST in near
 692                  * transactional manner.
 693                  */
 694                 rc = osp_sync_init(env, m);
 695                 if (rc)
 696                         GOTO(out_precreat, rc);
 697
 698                 rc = obd_fid_init(m->opd_obd, NULL, LUSTRE_SEQ_DATA);
 699                 if (rc) {
 700                         CERROR("%s: fid init error: rc = %d\n",
 701                                m->opd_obd->obd_name, rc);
 702                         GOTO(out, rc);
 703                 }
 704         }
 705         /*
 706          * Initiate connect to OST
 707          */
 708         ll_generate_random_uuid(uuid);
 709         class_uuid_unparse(uuid, &m->opd_cluuid);
 710
 711         imp = obd->u.cli.cl_import;
 712
 713         rc = ptlrpc_init_import(imp);
 714         if (rc)
 715                 GOTO(out, rc);
 716         if (osdname)
 717                 OBD_FREE(osdname, MAX_OBD_NAME);
 718         RETURN(0);
 719
 720 out:
 721         if (!m->opd_connect_mdt)
 722                 /* stop sync thread */
 723                 osp_sync_fini(m);
 724 out_precreat:
 725         /* stop precreate thread */
 726         if (!m->opd_connect_mdt)
 727                 osp_precreate_fini(m);
 728 out_last_used:
 729         osp_last_used_fini(env, m);
 730 out_proc:
 731         ptlrpc_lprocfs_unregister_obd(obd);
 732         lprocfs_obd_cleanup(obd);
 733         class_destroy_import(obd->u.cli.cl_import);
 734         client_obd_cleanup(obd);
 735 out_ref:
 736         ptlrpcd_decref();
 737 out_disconnect:
 738         if (m->opd_connect_mdt) {
 739                 struct client_obd *cli = &m->opd_obd->u.cli;
 740                 if (cli->cl_rpc_lock != NULL) {
 741                         OBD_FREE_PTR(cli->cl_rpc_lock);
 742                         cli->cl_rpc_lock = NULL;
 743                 }
 744         }
 745         obd_disconnect(m->opd_storage_exp);
 746 out_fini:
 747         if (osdname)
 748                 OBD_FREE(osdname, MAX_OBD_NAME);
 749         RETURN(rc);
 750 }
 751
 752 static struct lu_device *osp_device_free(const struct lu_env *env,
 753                                          struct lu_device *lu)
 754 {
 755         struct osp_device *m = lu2osp_dev(lu);
 756
 757         ENTRY;
 758
 759         if (cfs_atomic_read(&lu->ld_ref) && lu->ld_site) {
 760                 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
 761                 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
 762         }
 763         dt_device_fini(&m->opd_dt_dev);
 764         OBD_FREE_PTR(m);
 765         RETURN(NULL);
 766 }
 767
 768 static struct lu_device *osp_device_alloc(const struct lu_env *env,
 769                                           struct lu_device_type *t,
 770                                           struct lustre_cfg *lcfg)
 771 {
 772         struct osp_device *m;
 773         struct lu_device  *l;
 774
 775         OBD_ALLOC_PTR(m);
 776         if (m == NULL) {
 777                 l = ERR_PTR(-ENOMEM);
 778         } else {
 779                 int rc;
 780
 781                 l = osp2lu_dev(m);
 782                 dt_device_init(&m->opd_dt_dev, t);
 783                 rc = osp_init0(env, m, t, lcfg);
 784                 if (rc != 0) {
 785                         osp_device_free(env, l);
 786                         l = ERR_PTR(rc);
 787                 }
 788         }
 789         return l;
 790 }
 791
 792 static struct lu_device *osp_device_fini(const struct lu_env *env,
 793                                          struct lu_device *d)
 794 {
 795         struct osp_device *m = lu2osp_dev(d);
 796         struct obd_import *imp;
 797         int                rc;
 798
 799         ENTRY;
 800
 801         if (m->opd_storage_exp)
 802                 obd_disconnect(m->opd_storage_exp);
 803
 804         imp = m->opd_obd->u.cli.cl_import;
 805
 806         if (imp->imp_rq_pool) {
 807                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 808                 imp->imp_rq_pool = NULL;
 809         }
 810
 811         obd_cleanup_client_import(m->opd_obd);
 812
 813         if (m->opd_symlink)
 814                 lprocfs_remove(&m->opd_symlink);
 815
 816         LASSERT(m->opd_obd);
 817         ptlrpc_lprocfs_unregister_obd(m->opd_obd);
 818         lprocfs_obd_cleanup(m->opd_obd);
 819
 820         if (m->opd_connect_mdt) {
 821                 struct client_obd *cli = &m->opd_obd->u.cli;
 822                 if (cli->cl_rpc_lock != NULL) {
 823                         OBD_FREE_PTR(cli->cl_rpc_lock);
 824                         cli->cl_rpc_lock = NULL;
 825                 }
 826         }
 827
 828         rc = client_obd_cleanup(m->opd_obd);
 829         LASSERTF(rc == 0, "error %d\n", rc);
 830
 831         ptlrpcd_decref();
 832
 833         RETURN(NULL);
 834 }
 835
 836 static int osp_reconnect(const struct lu_env *env,
 837                          struct obd_export *exp, struct obd_device *obd,
 838                          struct obd_uuid *cluuid,
 839                          struct obd_connect_data *data,
 840                          void *localdata)
 841 {
 842         return 0;
 843 }
 844
 845 /*
 846  * we use exports to track all LOD users
 847  */
 848 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
 849                            struct obd_device *obd, struct obd_uuid *cluuid,
 850                            struct obd_connect_data *data, void *localdata)
 851 {
 852         struct osp_device       *osp = lu2osp_dev(obd->obd_lu_dev);
 853         struct obd_connect_data *ocd;
 854         struct obd_import       *imp;
 855         struct lustre_handle     conn;
 856         int                      rc;
 857
 858         ENTRY;
 859
 860         CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
 861
 862         rc = class_connect(&conn, obd, cluuid);
 863         if (rc)
 864                 RETURN(rc);
 865
 866         *exp = class_conn2export(&conn);
 867         /* Why should there ever be more than 1 connect? */
 868         osp->opd_connects++;
 869         LASSERT(osp->opd_connects == 1);
 870
 871         osp->opd_exp = *exp;
 872
 873         imp = osp->opd_obd->u.cli.cl_import;
 874         imp->imp_dlm_handle = conn;
 875
 876         LASSERT(data != NULL);
 877         LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
 878         ocd = &imp->imp_connect_data;
 879         *ocd = *data;
 880
 881         imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
 882
 883         ocd->ocd_version = LUSTRE_VERSION_CODE;
 884         ocd->ocd_index = data->ocd_index;
 885         imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
 886
 887         rc = ptlrpc_connect_import(imp);
 888         if (rc) {
 889                 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
 890                 GOTO(out, rc);
 891         }
 892
 893         ptlrpc_pinger_add_import(imp);
 894
 895         if (osp->opd_connect_mdt && data->ocd_index == 0) {
 896                 /* set seq controller export for MDC0 if exists */
 897                 struct seq_server_site *ss;
 898
 899                 ss = lu_site2seq(osp2lu_dev(osp)->ld_site);
 900                 ss->ss_control_exp = class_export_get(*exp);
 901                 ss->ss_server_fld->lsf_control_exp = *exp;
 902         }
 903 out:
 904         RETURN(rc);
 905 }
 906
 907 /*
 908  * once last export (we don't count self-export) disappeared
 909  * osp can be released
 910  */
 911 static int osp_obd_disconnect(struct obd_export *exp)
 912 {
 913         struct obd_device *obd = exp->exp_obd;
 914         struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
 915         int                rc;
 916         ENTRY;
 917
 918         /* Only disconnect the underlying layers on the final disconnect. */
 919         LASSERT(osp->opd_connects == 1);
 920         osp->opd_connects--;
 921
 922         rc = class_disconnect(exp);
 923         if (rc) {
 924                 CERROR("%s: class disconnect error: rc = %d\n",
 925                        obd->obd_name, rc);
 926                 RETURN(rc);
 927         }
 928
 929         /* destroy the device */
 930         class_manual_cleanup(obd);
 931
 932         RETURN(rc);
 933 }
 934
 935 /*
 936  * lprocfs helpers still use OBD API, let's keep obd_statfs() support
 937  */
 938 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
 939                           struct obd_statfs *osfs, __u64 max_age, __u32 flags)
 940 {
 941         struct obd_statfs       *msfs;
 942         struct ptlrpc_request   *req;
 943         struct obd_import       *imp = NULL;
 944         int                      rc;
 945
 946         ENTRY;
 947
 948         /* Since the request might also come from lprocfs, so we need
 949          * sync this with client_disconnect_export Bug15684 */
 950         down_read(&exp->exp_obd->u.cli.cl_sem);
 951         if (exp->exp_obd->u.cli.cl_import)
 952                 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
 953         up_read(&exp->exp_obd->u.cli.cl_sem);
 954         if (!imp)
 955                 RETURN(-ENODEV);
 956
 957         /* We could possibly pass max_age in the request (as an absolute
 958          * timestamp or a "seconds.usec ago") so the target can avoid doing
 959          * extra calls into the filesystem if that isn't necessary (e.g.
 960          * during mount that would help a bit).  Having relative timestamps
 961          * is not so great if request processing is slow, while absolute
 962          * timestamps are not ideal because they need time synchronization. */
 963         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
 964
 965         class_import_put(imp);
 966
 967         if (req == NULL)
 968                 RETURN(-ENOMEM);
 969
 970         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
 971         if (rc) {
 972                 ptlrpc_request_free(req);
 973                 RETURN(rc);
 974         }
 975         ptlrpc_request_set_replen(req);
 976         req->rq_request_portal = OST_CREATE_PORTAL;
 977         ptlrpc_at_set_req_timeout(req);
 978
 979         if (flags & OBD_STATFS_NODELAY) {
 980                 /* procfs requests not want stat in wait for avoid deadlock */
 981                 req->rq_no_resend = 1;
 982                 req->rq_no_delay = 1;
 983         }
 984
 985         rc = ptlrpc_queue_wait(req);
 986         if (rc)
 987                 GOTO(out, rc);
 988
 989         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
 990         if (msfs == NULL)
 991                 GOTO(out, rc = -EPROTO);
 992
 993         *osfs = *msfs;
 994
 995         EXIT;
 996 out:
 997         ptlrpc_req_finished(req);
 998         return rc;
 999 }
1000
1001 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
1002                             enum obd_import_event event)
1003 {
1004         struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1005
1006         switch (event) {
1007         case IMP_EVENT_DISCON:
1008                 d->opd_got_disconnected = 1;
1009                 d->opd_imp_connected = 0;
1010                 if (d->opd_connect_mdt)
1011                         break;
1012                 osp_pre_update_status(d, -ENODEV);
1013                 cfs_waitq_signal(&d->opd_pre_waitq);
1014                 CDEBUG(D_HA, "got disconnected\n");
1015                 break;
1016         case IMP_EVENT_INACTIVE:
1017                 d->opd_imp_active = 0;
1018                 if (d->opd_connect_mdt)
1019                         break;
1020                 osp_pre_update_status(d, -ENODEV);
1021                 cfs_waitq_signal(&d->opd_pre_waitq);
1022                 CDEBUG(D_HA, "got inactive\n");
1023                 break;
1024         case IMP_EVENT_ACTIVE:
1025                 d->opd_imp_active = 1;
1026                 if (d->opd_got_disconnected)
1027                         d->opd_new_connection = 1;
1028                 d->opd_imp_connected = 1;
1029                 d->opd_imp_seen_connected = 1;
1030                 if (d->opd_connect_mdt)
1031                         break;
1032                 cfs_waitq_signal(&d->opd_pre_waitq);
1033                 __osp_sync_check_for_work(d);
1034                 CDEBUG(D_HA, "got connected\n");
1035                 break;
1036         case IMP_EVENT_INVALIDATE:
1037                 if (obd->obd_namespace == NULL)
1038                         break;
1039                 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1040                 break;
1041         case IMP_EVENT_OCD:
1042         case IMP_EVENT_DEACTIVATE:
1043         case IMP_EVENT_ACTIVATE:
1044                 break;
1045         default:
1046                 CERROR("%s: unsupported import event: %#x\n",
1047                        obd->obd_name, event);
1048         }
1049         return 0;
1050 }
1051
1052 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1053                          void *karg, void *uarg)
1054 {
1055         struct obd_device       *obd = exp->exp_obd;
1056         struct osp_device       *d;
1057         struct obd_ioctl_data   *data = karg;
1058         int                      rc = 0;
1059
1060         ENTRY;
1061
1062         LASSERT(obd->obd_lu_dev);
1063         d = lu2osp_dev(obd->obd_lu_dev);
1064         LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1065
1066         if (!cfs_try_module_get(THIS_MODULE)) {
1067                 CERROR("%s: can't get module. Is it alive?", obd->obd_name);
1068                 return -EINVAL;
1069         }
1070
1071         switch (cmd) {
1072         case OBD_IOC_CLIENT_RECOVER:
1073                 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1074                                            data->ioc_inlbuf1, 0);
1075                 if (rc > 0)
1076                         rc = 0;
1077                 break;
1078         case IOC_OSC_SET_ACTIVE:
1079                 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1080                                               data->ioc_offset);
1081                 break;
1082         case OBD_IOC_PING_TARGET:
1083                 rc = ptlrpc_obd_ping(obd);
1084                 break;
1085         default:
1086                 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1087                        cmd, cfs_curproc_comm());
1088                 rc = -ENOTTY;
1089         }
1090         cfs_module_put(THIS_MODULE);
1091         return rc;
1092 }
1093
1094 static int osp_obd_health_check(const struct lu_env *env,
1095                                 struct obd_device *obd)
1096 {
1097         struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1098
1099         ENTRY;
1100
1101         /*
1102          * 1.8/2.0 behaviour is that OST being connected once at least
1103          * is considired "healthy". and one "healty" OST is enough to
1104          * allow lustre clients to connect to MDS
1105          */
1106         LASSERT(d);
1107         RETURN(!d->opd_imp_seen_connected);
1108 }
1109
1110 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1111 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1112 static void osp_key_exit(const struct lu_context *ctx,
1113                          struct lu_context_key *key, void *data)
1114 {
1115         struct osp_thread_info *info = data;
1116
1117         info->osi_attr.la_valid = 0;
1118 }
1119
1120 struct lu_context_key osp_thread_key = {
1121         .lct_tags = LCT_MD_THREAD,
1122         .lct_init = osp_key_init,
1123         .lct_fini = osp_key_fini,
1124         .lct_exit = osp_key_exit
1125 };
1126
1127 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1128 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1129
1130 struct lu_context_key osp_txn_key = {
1131         .lct_tags = LCT_OSP_THREAD,
1132         .lct_init = osp_txn_key_init,
1133         .lct_fini = osp_txn_key_fini
1134 };
1135 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1136
1137 static struct lu_device_type_operations osp_device_type_ops = {
1138         .ldto_init           = osp_type_init,
1139         .ldto_fini           = osp_type_fini,
1140
1141         .ldto_start          = osp_type_start,
1142         .ldto_stop           = osp_type_stop,
1143
1144         .ldto_device_alloc   = osp_device_alloc,
1145         .ldto_device_free    = osp_device_free,
1146
1147         .ldto_device_fini    = osp_device_fini
1148 };
1149
1150 static struct lu_device_type osp_device_type = {
1151         .ldt_tags     = LU_DEVICE_DT,
1152         .ldt_name     = LUSTRE_OSP_NAME,
1153         .ldt_ops      = &osp_device_type_ops,
1154         .ldt_ctx_tags = LCT_MD_THREAD
1155 };
1156
1157 static struct obd_ops osp_obd_device_ops = {
1158         .o_owner        = THIS_MODULE,
1159         .o_add_conn     = client_import_add_conn,
1160         .o_del_conn     = client_import_del_conn,
1161         .o_reconnect    = osp_reconnect,
1162         .o_connect      = osp_obd_connect,
1163         .o_disconnect   = osp_obd_disconnect,
1164         .o_health_check = osp_obd_health_check,
1165         .o_import_event = osp_import_event,
1166         .o_iocontrol    = osp_iocontrol,
1167         .o_statfs       = osp_obd_statfs,
1168         .o_fid_init     = client_fid_init,
1169         .o_fid_fini     = client_fid_fini,
1170 };
1171
1172 struct llog_operations osp_mds_ost_orig_logops;
1173
1174 static int __init osp_mod_init(void)
1175 {
1176         struct lprocfs_static_vars       lvars;
1177         cfs_proc_dir_entry_t            *osc_proc_dir;
1178         int                              rc;
1179
1180         rc = lu_kmem_init(osp_caches);
1181         if (rc)
1182                 return rc;
1183
1184         lprocfs_osp_init_vars(&lvars);
1185
1186         rc = class_register_type(&osp_obd_device_ops, NULL, lvars.module_vars,
1187                                  LUSTRE_OSP_NAME, &osp_device_type);
1188
1189         /* create "osc" entry in procfs for compatibility purposes */
1190         if (rc != 0) {
1191                 lu_kmem_fini(osp_caches);
1192                 return rc;
1193         }
1194
1195         lprocfs_lwp_init_vars(&lvars);
1196
1197         rc = class_register_type(&lwp_obd_device_ops, NULL, lvars.module_vars,
1198                                  LUSTRE_LWP_NAME, &lwp_device_type);
1199         if (rc != 0) {
1200                 class_unregister_type(LUSTRE_OSP_NAME);
1201                 lu_kmem_fini(osp_caches);
1202                 return rc;
1203         }
1204
1205         /* Note: add_rec/delcare_add_rec will be only used by catalogs */
1206         osp_mds_ost_orig_logops = llog_osd_ops;
1207         osp_mds_ost_orig_logops.lop_add = llog_cat_add_rec;
1208         osp_mds_ost_orig_logops.lop_declare_add = llog_cat_declare_add_rec;
1209
1210         osc_proc_dir = lprocfs_srch(proc_lustre_root, "osc");
1211         if (osc_proc_dir == NULL) {
1212                 osc_proc_dir = lprocfs_register("osc", proc_lustre_root, NULL,
1213                                                 NULL);
1214                 if (IS_ERR(osc_proc_dir))
1215                         CERROR("osp: can't create compat entry \"osc\": %d\n",
1216                                (int) PTR_ERR(osc_proc_dir));
1217         }
1218         return rc;
1219 }
1220
1221 static void __exit osp_mod_exit(void)
1222 {
1223         lprocfs_try_remove_proc_entry("osc", proc_lustre_root);
1224
1225         class_unregister_type(LUSTRE_LWP_NAME);
1226         class_unregister_type(LUSTRE_OSP_NAME);
1227         lu_kmem_fini(osp_caches);
1228 }
1229
1230 MODULE_AUTHOR("Intel, Inc. <http://www.intel.com/>");
1231 MODULE_DESCRIPTION("Lustre OST Proxy Device ("LUSTRE_OSP_NAME")");
1232 MODULE_LICENSE("GPL");
1233
1234 cfs_module(osp, LUSTRE_VERSION_STRING, osp_mod_init, osp_mod_exit);