Whamcloud - gitweb
7f3311cf06f3fdb396bb9cd613fa040759b4143c
[fs/lustre-release.git] / 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, 2013, 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 #define DEBUG_SUBSYSTEM S_MDS
46
47 #include <obd_class.h>
48 #include <lustre_param.h>
49 #include <lustre_log.h>
50 #include <lustre_mdc.h>
51
52 #include "osp_internal.h"
53
54 /* Slab for OSP object allocation */
55 struct kmem_cache *osp_object_kmem;
56
57 static struct lu_kmem_descr osp_caches[] = {
58         {
59                 .ckd_cache = &osp_object_kmem,
60                 .ckd_name  = "osp_obj",
61                 .ckd_size  = sizeof(struct osp_object)
62         },
63         {
64                 .ckd_cache = NULL
65         }
66 };
67
68 struct lu_object *osp_object_alloc(const struct lu_env *env,
69                                    const struct lu_object_header *hdr,
70                                    struct lu_device *d)
71 {
72         struct lu_object_header *h = NULL;
73         struct osp_object       *o;
74         struct lu_object        *l;
75
76         OBD_SLAB_ALLOC_PTR_GFP(o, osp_object_kmem, __GFP_IO);
77         if (o != NULL) {
78                 l = &o->opo_obj.do_lu;
79
80                 /* If hdr is NULL, it means the object is not built
81                  * from the top dev(MDT/OST), usually it happens when
82                  * building striped object, like data object on MDT or
83                  * striped object for directory */
84                 if (hdr == NULL) {
85                         h = &o->opo_header;
86                         lu_object_header_init(h);
87                         dt_object_init(&o->opo_obj, h, d);
88                         lu_object_add_top(h, l);
89                 } else {
90                         dt_object_init(&o->opo_obj, h, d);
91                 }
92
93                 l->lo_ops = &osp_lu_obj_ops;
94
95                 return l;
96         } else {
97                 return NULL;
98         }
99 }
100
101 static struct dt_object
102 *osp_find_or_create(const struct lu_env *env, struct osp_device *osp,
103                     struct lu_attr *attr, __u32 reg_id)
104 {
105         struct osp_thread_info *osi = osp_env_info(env);
106         struct dt_object_format dof = { 0 };
107         struct dt_object       *dto;
108         int                  rc;
109         ENTRY;
110
111         lu_local_obj_fid(&osi->osi_fid, reg_id);
112         attr->la_valid = LA_MODE;
113         attr->la_mode = S_IFREG | 0644;
114         dof.dof_type = DFT_REGULAR;
115         dto = dt_find_or_create(env, osp->opd_storage, &osi->osi_fid,
116                                 &dof, attr);
117         if (IS_ERR(dto))
118                 RETURN(dto);
119
120         rc = dt_attr_get(env, dto, attr, NULL);
121         if (rc) {
122                 CERROR("%s: can't be initialized: rc = %d\n",
123                        osp->opd_obd->obd_name, rc);
124                 lu_object_put(env, &dto->do_lu);
125                 RETURN(ERR_PTR(rc));
126         }
127         RETURN(dto);
128 }
129
130 static int osp_write_local_file(const struct lu_env *env,
131                                 struct osp_device *osp,
132                                 struct dt_object *dt_obj,
133                                 struct lu_buf *buf,
134                                 loff_t offset)
135 {
136         struct thandle *th;
137         int rc;
138
139         th = dt_trans_create(env, osp->opd_storage);
140         if (IS_ERR(th))
141                 RETURN(PTR_ERR(th));
142
143         rc = dt_declare_record_write(env, dt_obj, buf->lb_len, offset, th);
144         if (rc)
145                 GOTO(out, rc);
146         rc = dt_trans_start_local(env, osp->opd_storage, th);
147         if (rc)
148                 GOTO(out, rc);
149
150         rc = dt_record_write(env, dt_obj, buf, &offset, th);
151 out:
152         dt_trans_stop(env, osp->opd_storage, th);
153         RETURN(rc);
154 }
155
156 static int osp_init_last_objid(const struct lu_env *env, struct osp_device *osp)
157 {
158         struct osp_thread_info  *osi = osp_env_info(env);
159         struct lu_fid           *fid = &osp->opd_last_used_fid;
160         struct dt_object        *dto;
161         int                     rc;
162         ENTRY;
163
164         dto = osp_find_or_create(env, osp, &osi->osi_attr, MDD_LOV_OBJ_OID);
165         if (IS_ERR(dto))
166                 RETURN(PTR_ERR(dto));
167         /* object will be released in device cleanup path */
168         if (osi->osi_attr.la_size >=
169             sizeof(osi->osi_id) * (osp->opd_index + 1)) {
170                 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
171                                    osp->opd_index);
172                 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
173                 if (rc != 0)
174                         GOTO(out, rc);
175         } else {
176                 fid->f_oid = 0;
177                 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
178                                    osp->opd_index);
179                 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
180                                           osi->osi_off);
181         }
182         osp->opd_last_used_oid_file = dto;
183         RETURN(0);
184 out:
185         /* object will be released in device cleanup path */
186         CERROR("%s: can't initialize lov_objid: rc = %d\n",
187                osp->opd_obd->obd_name, rc);
188         lu_object_put(env, &dto->do_lu);
189         osp->opd_last_used_oid_file = NULL;
190         RETURN(rc);
191 }
192
193 static int osp_init_last_seq(const struct lu_env *env, struct osp_device *osp)
194 {
195         struct osp_thread_info  *osi = osp_env_info(env);
196         struct lu_fid           *fid = &osp->opd_last_used_fid;
197         struct dt_object        *dto;
198         int                     rc;
199         ENTRY;
200
201         dto = osp_find_or_create(env, osp, &osi->osi_attr, MDD_LOV_OBJ_OSEQ);
202         if (IS_ERR(dto))
203                 RETURN(PTR_ERR(dto));
204
205         /* object will be released in device cleanup path */
206         if (osi->osi_attr.la_size >=
207             sizeof(osi->osi_id) * (osp->opd_index + 1)) {
208                 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
209                                    osp->opd_index);
210                 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
211                 if (rc != 0)
212                         GOTO(out, rc);
213         } else {
214                 fid->f_seq = 0;
215                 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
216                                     osp->opd_index);
217                 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
218                                           osi->osi_off);
219         }
220         osp->opd_last_used_seq_file = dto;
221         RETURN(0);
222 out:
223         /* object will be released in device cleanup path */
224         CERROR("%s: can't initialize lov_seq: rc = %d\n",
225                osp->opd_obd->obd_name, rc);
226         lu_object_put(env, &dto->do_lu);
227         osp->opd_last_used_seq_file = NULL;
228         RETURN(rc);
229 }
230
231 static int osp_last_used_init(const struct lu_env *env, struct osp_device *osp)
232 {
233         struct osp_thread_info *osi = osp_env_info(env);
234         int                  rc;
235         ENTRY;
236
237         fid_zero(&osp->opd_last_used_fid);
238         rc = osp_init_last_objid(env, osp);
239         if (rc < 0) {
240                 CERROR("%s: Can not get ids %d from old objid!\n",
241                        osp->opd_obd->obd_name, rc);
242                 RETURN(rc);
243         }
244
245         rc = osp_init_last_seq(env, osp);
246         if (rc < 0) {
247                 CERROR("%s: Can not get ids %d from old objid!\n",
248                        osp->opd_obd->obd_name, rc);
249                 GOTO(out, rc);
250         }
251
252         if (fid_oid(&osp->opd_last_used_fid) != 0 &&
253             fid_seq(&osp->opd_last_used_fid) == 0) {
254                 /* Just upgrade from the old version,
255                  * set the seq to be IDIF */
256                 osp->opd_last_used_fid.f_seq =
257                    fid_idif_seq(fid_oid(&osp->opd_last_used_fid),
258                                 osp->opd_index);
259                 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off,
260                                     &osp->opd_last_used_fid.f_seq,
261                                     osp->opd_index);
262                 rc = osp_write_local_file(env, osp, osp->opd_last_used_seq_file,
263                                           &osi->osi_lb, osi->osi_off);
264                 if (rc) {
265                         CERROR("%s : Can not write seq file: rc = %d\n",
266                                osp->opd_obd->obd_name, rc);
267                         GOTO(out, rc);
268                 }
269         }
270
271         if (!fid_is_zero(&osp->opd_last_used_fid) &&
272                  !fid_is_sane(&osp->opd_last_used_fid)) {
273                 CERROR("%s: Got invalid FID "DFID"\n", osp->opd_obd->obd_name,
274                         PFID(&osp->opd_last_used_fid));
275                 GOTO(out, rc = -EINVAL);
276         }
277
278         CDEBUG(D_INFO, "%s: Init last used fid "DFID"\n",
279                osp->opd_obd->obd_name, PFID(&osp->opd_last_used_fid));
280 out:
281         if (rc != 0) {
282                 if (osp->opd_last_used_oid_file != NULL) {
283                         lu_object_put(env, &osp->opd_last_used_oid_file->do_lu);
284                         osp->opd_last_used_oid_file = NULL;
285                 }
286                 if (osp->opd_last_used_seq_file != NULL) {
287                         lu_object_put(env, &osp->opd_last_used_seq_file->do_lu);
288                         osp->opd_last_used_seq_file = NULL;
289                 }
290         }
291
292         RETURN(rc);
293 }
294
295 static void osp_last_used_fini(const struct lu_env *env, struct osp_device *d)
296 {
297         /* release last_used file */
298         if (d->opd_last_used_oid_file != NULL) {
299                 lu_object_put(env, &d->opd_last_used_oid_file->do_lu);
300                 d->opd_last_used_oid_file = NULL;
301         }
302
303         if (d->opd_last_used_seq_file != NULL) {
304                 lu_object_put(env, &d->opd_last_used_seq_file->do_lu);
305                 d->opd_last_used_seq_file = NULL;
306         }
307 }
308
309 static int osp_disconnect(struct osp_device *d)
310 {
311         struct obd_import *imp;
312         int rc = 0;
313
314         imp = d->opd_obd->u.cli.cl_import;
315
316         /* Mark import deactivated now, so we don't try to reconnect if any
317          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
318          * fully deactivate the import, or that would drop all requests. */
319         LASSERT(imp != NULL);
320         spin_lock(&imp->imp_lock);
321         imp->imp_deactive = 1;
322         spin_unlock(&imp->imp_lock);
323
324         ptlrpc_deactivate_import(imp);
325
326         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
327          * delete it regardless.  (It's safe to delete an import that was
328          * never added.) */
329         (void)ptlrpc_pinger_del_import(imp);
330
331         rc = ptlrpc_disconnect_import(imp, 0);
332         if (rc != 0)
333                 CERROR("%s: can't disconnect: rc = %d\n",
334                        d->opd_obd->obd_name, rc);
335
336         ptlrpc_invalidate_import(imp);
337
338         RETURN(rc);
339 }
340
341 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
342 {
343         int                      rc = 0;
344         ENTRY;
345
346         LASSERT(env);
347
348         rc = osp_disconnect(d);
349
350         osp_sync_fini(d);
351
352         if (!d->opd_connect_mdt) {
353                 /* stop precreate thread */
354                 osp_precreate_fini(d);
355
356                 /* release last_used file */
357                 osp_last_used_fini(env, d);
358         }
359
360         obd_fid_fini(d->opd_obd);
361
362         RETURN(rc);
363 }
364
365 static int osp_process_config(const struct lu_env *env,
366                               struct lu_device *dev, struct lustre_cfg *lcfg)
367 {
368         struct osp_device               *d = lu2osp_dev(dev);
369         struct lprocfs_static_vars       lvars = { 0 };
370         int                              rc;
371
372         ENTRY;
373
374         switch (lcfg->lcfg_command) {
375         case LCFG_PRE_CLEANUP:
376                 rc = osp_disconnect(d);
377                 break;
378         case LCFG_CLEANUP:
379                 lu_dev_del_linkage(dev->ld_site, dev);
380                 rc = osp_shutdown(env, d);
381                 break;
382         case LCFG_PARAM:
383                 lprocfs_osp_init_vars(&lvars);
384
385                 LASSERT(d->opd_obd);
386                 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
387                                               lcfg, d->opd_obd);
388                 if (rc > 0)
389                         rc = 0;
390                 if (rc == -ENOSYS) {
391                         /* class_process_proc_param() haven't found matching
392                          * parameter and returned ENOSYS so that layer(s)
393                          * below could use that. But OSP is the bottom, so
394                          * just ignore it */
395                         CERROR("%s: unknown param %s\n",
396                                (char *)lustre_cfg_string(lcfg, 0),
397                                (char *)lustre_cfg_string(lcfg, 1));
398                         rc = 0;
399                 }
400                 break;
401         default:
402                 CERROR("%s: unknown command %u\n",
403                        (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
404                 rc = 0;
405                 break;
406         }
407
408         RETURN(rc);
409 }
410
411 static int osp_recovery_complete(const struct lu_env *env,
412                                  struct lu_device *dev)
413 {
414         struct osp_device       *osp = lu2osp_dev(dev);
415         int                      rc = 0;
416
417         ENTRY;
418         osp->opd_recovery_completed = 1;
419
420         if (!osp->opd_connect_mdt && osp->opd_pre != NULL)
421                 wake_up(&osp->opd_pre_waitq);
422         RETURN(rc);
423 }
424
425 const struct lu_device_operations osp_lu_ops = {
426         .ldo_object_alloc       = osp_object_alloc,
427         .ldo_process_config     = osp_process_config,
428         .ldo_recovery_complete  = osp_recovery_complete,
429 };
430
431 /**
432  * provides with statfs from corresponded OST
433  *
434  */
435 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
436                       struct obd_statfs *sfs)
437 {
438         struct osp_device *d = dt2osp_dev(dev);
439
440         ENTRY;
441
442         if (unlikely(d->opd_imp_active == 0))
443                 RETURN(-ENOTCONN);
444
445         if (d->opd_pre == NULL)
446                 RETURN(0);
447
448         /* return recently updated data */
449         *sfs = d->opd_statfs;
450
451         /*
452          * layer above osp (usually lod) can use ffree to estimate
453          * how many objects are available for immediate creation
454          */
455         spin_lock(&d->opd_pre_lock);
456         LASSERTF(fid_seq(&d->opd_pre_last_created_fid) ==
457                  fid_seq(&d->opd_pre_used_fid),
458                  "last_created "DFID", next_fid "DFID"\n",
459                  PFID(&d->opd_pre_last_created_fid),
460                  PFID(&d->opd_pre_used_fid));
461         sfs->os_fprecreated = fid_oid(&d->opd_pre_last_created_fid) -
462                               fid_oid(&d->opd_pre_used_fid);
463         sfs->os_fprecreated -= d->opd_pre_reserved;
464         spin_unlock(&d->opd_pre_lock);
465
466         LASSERT(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2);
467
468         CDEBUG(D_OTHER, "%s: "LPU64" blocks, "LPU64" free, "LPU64" avail, "
469                LPU64" files, "LPU64" free files\n", d->opd_obd->obd_name,
470                sfs->os_blocks, sfs->os_bfree, sfs->os_bavail,
471                sfs->os_files, sfs->os_ffree);
472         RETURN(0);
473 }
474
475 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
476 {
477         ENTRY;
478
479         /*
480          * XXX: wake up sync thread, command it to start flushing asap?
481          */
482
483         RETURN(0);
484 }
485
486 const struct dt_device_operations osp_dt_ops = {
487         .dt_statfs       = osp_statfs,
488         .dt_sync         = osp_sync,
489         .dt_trans_create = osp_trans_create,
490         .dt_trans_start  = osp_trans_start,
491         .dt_trans_stop   = osp_trans_stop,
492 };
493
494 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *m,
495                               const char *nextdev)
496 {
497         struct obd_connect_data *data = NULL;
498         struct obd_device       *obd;
499         int                      rc;
500
501         ENTRY;
502
503         LASSERT(m->opd_storage_exp == NULL);
504
505         OBD_ALLOC_PTR(data);
506         if (data == NULL)
507                 RETURN(-ENOMEM);
508
509         obd = class_name2obd(nextdev);
510         if (obd == NULL) {
511                 CERROR("%s: can't locate next device: %s\n",
512                        m->opd_obd->obd_name, nextdev);
513                 GOTO(out, rc = -ENOTCONN);
514         }
515
516         rc = obd_connect(env, &m->opd_storage_exp, obd, &obd->obd_uuid, data,
517                          NULL);
518         if (rc) {
519                 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
520                        m->opd_obd->obd_name, nextdev, rc);
521                 GOTO(out, rc);
522         }
523
524         m->opd_dt_dev.dd_lu_dev.ld_site =
525                 m->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
526         LASSERT(m->opd_dt_dev.dd_lu_dev.ld_site);
527         m->opd_storage = lu2dt_dev(m->opd_storage_exp->exp_obd->obd_lu_dev);
528
529 out:
530         OBD_FREE_PTR(data);
531         RETURN(rc);
532 }
533
534 static int osp_init0(const struct lu_env *env, struct osp_device *m,
535                      struct lu_device_type *ldt, struct lustre_cfg *cfg)
536 {
537         struct obd_device       *obd;
538         struct obd_import       *imp;
539         class_uuid_t            uuid;
540         char                    *src, *tgt, *mdt, *osdname = NULL;
541         int                     rc;
542         long                    idx;
543
544         ENTRY;
545
546         mutex_init(&m->opd_async_requests_mutex);
547         obd = class_name2obd(lustre_cfg_string(cfg, 0));
548         if (obd == NULL) {
549                 CERROR("Cannot find obd with name %s\n",
550                        lustre_cfg_string(cfg, 0));
551                 RETURN(-ENODEV);
552         }
553         m->opd_obd = obd;
554
555         /* There is no record in the MDT configuration for the local disk
556          * device, so we have to extract this from elsewhere in the profile.
557          * The only information we get at setup is from the OSC records:
558          * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
559          * Note that 1.8 generated configs are missing the -MDTxxxx part.
560          * We need to reconstruct the name of the underlying OSD from this:
561          * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".  We
562          * also need to determine the OST index from this - will be used
563          * to calculate the offset in shared lov_objids file later */
564
565         src = lustre_cfg_string(cfg, 0);
566         if (src == NULL)
567                 RETURN(-EINVAL);
568
569         tgt = strrchr(src, '-');
570         if (tgt == NULL) {
571                 CERROR("%s: invalid target name %s\n",
572                        m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
573                 RETURN(-EINVAL);
574         }
575
576         if (strncmp(tgt, "-osc", 4) == 0) {
577                 /* Old OSC name fsname-OSTXXXX-osc */
578                 for (tgt--; tgt > src && *tgt != '-'; tgt--)
579                         ;
580                 if (tgt == src) {
581                         CERROR("%s: invalid target name %s\n",
582                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
583                         RETURN(-EINVAL);
584                 }
585
586                 if (strncmp(tgt, "-OST", 4) != 0) {
587                         CERROR("%s: invalid target name %s\n",
588                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
589                         RETURN(-EINVAL);
590                 }
591
592                 idx = simple_strtol(tgt + 4, &mdt, 16);
593                 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
594                         CERROR("%s: invalid OST index in '%s'\n",
595                                m->opd_obd->obd_name, src);
596                         RETURN(-EINVAL);
597                 }
598                 m->opd_index = idx;
599                 m->opd_group = 0;
600                 idx = tgt - src;
601         } else {
602                 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
603                 if (strncmp(tgt, "-MDT", 4) != 0 &&
604                     strncmp(tgt, "-OST", 4) != 0) {
605                         CERROR("%s: invalid target name %s\n",
606                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
607                         RETURN(-EINVAL);
608                 }
609
610                 idx = simple_strtol(tgt + 4, &mdt, 16);
611                 if (*mdt != '\0' || idx > INT_MAX || idx < 0) {
612                         CERROR("%s: invalid OST index in '%s'\n",
613                                m->opd_obd->obd_name, src);
614                         RETURN(-EINVAL);
615                 }
616
617                 /* Get MDT index from the name and set it to opd_group,
618                  * which will be used by OSP to connect with OST */
619                 m->opd_group = idx;
620                 if (tgt - src <= 12) {
621                         CERROR("%s: invalid mdt index retrieve from %s\n",
622                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
623                         RETURN(-EINVAL);
624                 }
625
626                 if (strncmp(tgt - 12, "-MDT", 4) == 0)
627                         m->opd_connect_mdt = 1;
628
629                 idx = simple_strtol(tgt - 8, &mdt, 16);
630                 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
631                         CERROR("%s: invalid OST index in '%s'\n",
632                                m->opd_obd->obd_name, src);
633                         RETURN(-EINVAL);
634                 }
635
636                 m->opd_index = idx;
637                 idx = tgt - src - 12;
638         }
639         /* check the fsname length, and after this everything else will fit */
640         if (idx > MTI_NAME_MAXLEN) {
641                 CERROR("%s: fsname too long in '%s'\n",
642                        m->opd_obd->obd_name, src);
643                 RETURN(-EINVAL);
644         }
645
646         OBD_ALLOC(osdname, MAX_OBD_NAME);
647         if (osdname == NULL)
648                 RETURN(-ENOMEM);
649
650         memcpy(osdname, src, idx); /* copy just the fsname part */
651         osdname[idx] = '\0';
652
653         mdt = strstr(mdt, "-MDT");
654         if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
655                 strcat(osdname, "-MDT0000");
656         else
657                 strcat(osdname, mdt);
658         strcat(osdname, "-osd");
659         CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
660
661         if (m->opd_connect_mdt) {
662                 struct client_obd *cli = &m->opd_obd->u.cli;
663
664                 OBD_ALLOC(cli->cl_rpc_lock, sizeof(*cli->cl_rpc_lock));
665                 if (!cli->cl_rpc_lock)
666                         GOTO(out_fini, rc = -ENOMEM);
667                 osp_init_rpc_lock(cli->cl_rpc_lock);
668         }
669
670         m->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
671         m->opd_dt_dev.dd_ops = &osp_dt_ops;
672         obd->obd_lu_dev = &m->opd_dt_dev.dd_lu_dev;
673
674         rc = osp_connect_to_osd(env, m, osdname);
675         if (rc)
676                 GOTO(out_fini, rc);
677
678         rc = ptlrpcd_addref();
679         if (rc)
680                 GOTO(out_disconnect, rc);
681
682         rc = client_obd_setup(obd, cfg);
683         if (rc) {
684                 CERROR("%s: can't setup obd: %d\n", m->opd_obd->obd_name, rc);
685                 GOTO(out_ref, rc);
686         }
687
688         osp_lprocfs_init(m);
689
690         rc = obd_fid_init(m->opd_obd, NULL, m->opd_connect_mdt ?
691                           LUSTRE_SEQ_METADATA : LUSTRE_SEQ_DATA);
692         if (rc) {
693                 CERROR("%s: fid init error: rc = %d\n",
694                        m->opd_obd->obd_name, rc);
695                 GOTO(out_proc, rc);
696         }
697
698         if (!m->opd_connect_mdt) {
699                 /* Initialize last id from the storage - will be
700                  * used in orphan cleanup. */
701                 rc = osp_last_used_init(env, m);
702                 if (rc)
703                         GOTO(out_proc, rc);
704
705
706                 /* Initialize precreation thread, it handles new
707                  * connections as well. */
708                 rc = osp_init_precreate(m);
709                 if (rc)
710                         GOTO(out_last_used, rc);
711         }
712
713         /*
714          * Initialize synhronization mechanism taking
715          * care of propogating changes to OST in near
716          * transactional manner.
717          */
718         rc = osp_sync_init(env, m);
719         if (rc)
720                 GOTO(out_precreat, rc);
721
722         /*
723          * Initiate connect to OST
724          */
725         ll_generate_random_uuid(uuid);
726         class_uuid_unparse(uuid, &m->opd_cluuid);
727
728         imp = obd->u.cli.cl_import;
729
730         rc = ptlrpc_init_import(imp);
731         if (rc)
732                 GOTO(out, rc);
733         if (osdname)
734                 OBD_FREE(osdname, MAX_OBD_NAME);
735         RETURN(0);
736
737 out:
738         /* stop sync thread */
739         osp_sync_fini(m);
740 out_precreat:
741         /* stop precreate thread */
742         if (!m->opd_connect_mdt)
743                 osp_precreate_fini(m);
744 out_last_used:
745         if (!m->opd_connect_mdt)
746                 osp_last_used_fini(env, m);
747 out_proc:
748         ptlrpc_lprocfs_unregister_obd(obd);
749         lprocfs_obd_cleanup(obd);
750         if (m->opd_symlink)
751                 lprocfs_remove(&m->opd_symlink);
752         client_obd_cleanup(obd);
753 out_ref:
754         ptlrpcd_decref();
755 out_disconnect:
756         if (m->opd_connect_mdt) {
757                 struct client_obd *cli = &m->opd_obd->u.cli;
758                 if (cli->cl_rpc_lock != NULL) {
759                         OBD_FREE_PTR(cli->cl_rpc_lock);
760                         cli->cl_rpc_lock = NULL;
761                 }
762         }
763         obd_disconnect(m->opd_storage_exp);
764 out_fini:
765         if (osdname)
766                 OBD_FREE(osdname, MAX_OBD_NAME);
767         RETURN(rc);
768 }
769
770 static struct lu_device *osp_device_free(const struct lu_env *env,
771                                          struct lu_device *lu)
772 {
773         struct osp_device *m = lu2osp_dev(lu);
774
775         ENTRY;
776
777         if (cfs_atomic_read(&lu->ld_ref) && lu->ld_site) {
778                 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
779                 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
780         }
781         dt_device_fini(&m->opd_dt_dev);
782         OBD_FREE_PTR(m);
783         RETURN(NULL);
784 }
785
786 static struct lu_device *osp_device_alloc(const struct lu_env *env,
787                                           struct lu_device_type *t,
788                                           struct lustre_cfg *lcfg)
789 {
790         struct osp_device *m;
791         struct lu_device  *l;
792
793         OBD_ALLOC_PTR(m);
794         if (m == NULL) {
795                 l = ERR_PTR(-ENOMEM);
796         } else {
797                 int rc;
798
799                 l = osp2lu_dev(m);
800                 dt_device_init(&m->opd_dt_dev, t);
801                 rc = osp_init0(env, m, t, lcfg);
802                 if (rc != 0) {
803                         osp_device_free(env, l);
804                         l = ERR_PTR(rc);
805                 }
806         }
807         return l;
808 }
809
810 static struct lu_device *osp_device_fini(const struct lu_env *env,
811                                          struct lu_device *d)
812 {
813         struct osp_device *m = lu2osp_dev(d);
814         struct obd_import *imp;
815         int                rc;
816
817         ENTRY;
818
819         if (m->opd_async_requests != NULL) {
820                 out_destroy_update_req(m->opd_async_requests);
821                 m->opd_async_requests = NULL;
822         }
823
824         if (m->opd_storage_exp)
825                 obd_disconnect(m->opd_storage_exp);
826
827         imp = m->opd_obd->u.cli.cl_import;
828
829         if (imp->imp_rq_pool) {
830                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
831                 imp->imp_rq_pool = NULL;
832         }
833
834         if (m->opd_symlink)
835                 lprocfs_remove(&m->opd_symlink);
836
837         LASSERT(m->opd_obd);
838         ptlrpc_lprocfs_unregister_obd(m->opd_obd);
839         lprocfs_obd_cleanup(m->opd_obd);
840
841         if (m->opd_connect_mdt) {
842                 struct client_obd *cli = &m->opd_obd->u.cli;
843                 if (cli->cl_rpc_lock != NULL) {
844                         OBD_FREE_PTR(cli->cl_rpc_lock);
845                         cli->cl_rpc_lock = NULL;
846                 }
847         }
848
849         rc = client_obd_cleanup(m->opd_obd);
850         LASSERTF(rc == 0, "error %d\n", rc);
851
852         ptlrpcd_decref();
853
854         RETURN(NULL);
855 }
856
857 static int osp_reconnect(const struct lu_env *env,
858                          struct obd_export *exp, struct obd_device *obd,
859                          struct obd_uuid *cluuid,
860                          struct obd_connect_data *data,
861                          void *localdata)
862 {
863         return 0;
864 }
865
866 static int osp_prepare_fid_client(struct osp_device *osp)
867 {
868         LASSERT(osp->opd_obd->u.cli.cl_seq != NULL);
869         if (osp->opd_obd->u.cli.cl_seq->lcs_exp != NULL)
870                 return 0;
871
872         LASSERT(osp->opd_exp != NULL);
873         osp->opd_obd->u.cli.cl_seq->lcs_exp =
874                                 class_export_get(osp->opd_exp);
875         return 0;
876 }
877
878 /*
879  * we use exports to track all LOD users
880  */
881 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
882                            struct obd_device *obd, struct obd_uuid *cluuid,
883                            struct obd_connect_data *data, void *localdata)
884 {
885         struct osp_device       *osp = lu2osp_dev(obd->obd_lu_dev);
886         struct obd_connect_data *ocd;
887         struct obd_import       *imp;
888         struct lustre_handle     conn;
889         int                      rc;
890
891         ENTRY;
892
893         CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
894
895         rc = class_connect(&conn, obd, cluuid);
896         if (rc)
897                 RETURN(rc);
898
899         *exp = class_conn2export(&conn);
900         /* Why should there ever be more than 1 connect? */
901         osp->opd_connects++;
902         LASSERT(osp->opd_connects == 1);
903
904         osp->opd_exp = *exp;
905
906         imp = osp->opd_obd->u.cli.cl_import;
907         imp->imp_dlm_handle = conn;
908
909         LASSERT(data != NULL);
910         LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
911         ocd = &imp->imp_connect_data;
912         *ocd = *data;
913
914         imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
915
916         ocd->ocd_version = LUSTRE_VERSION_CODE;
917         ocd->ocd_index = data->ocd_index;
918         imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
919
920         rc = ptlrpc_connect_import(imp);
921         if (rc) {
922                 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
923                 GOTO(out, rc);
924         }
925
926         ptlrpc_pinger_add_import(imp);
927 out:
928         RETURN(rc);
929 }
930
931 /*
932  * once last export (we don't count self-export) disappeared
933  * osp can be released
934  */
935 static int osp_obd_disconnect(struct obd_export *exp)
936 {
937         struct obd_device *obd = exp->exp_obd;
938         struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
939         int                rc;
940         ENTRY;
941
942         /* Only disconnect the underlying layers on the final disconnect. */
943         LASSERT(osp->opd_connects == 1);
944         osp->opd_connects--;
945
946         rc = class_disconnect(exp);
947         if (rc) {
948                 CERROR("%s: class disconnect error: rc = %d\n",
949                        obd->obd_name, rc);
950                 RETURN(rc);
951         }
952
953         /* destroy the device */
954         class_manual_cleanup(obd);
955
956         RETURN(rc);
957 }
958
959 /*
960  * lprocfs helpers still use OBD API, let's keep obd_statfs() support
961  */
962 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
963                           struct obd_statfs *osfs, __u64 max_age, __u32 flags)
964 {
965         struct obd_statfs       *msfs;
966         struct ptlrpc_request   *req;
967         struct obd_import       *imp = NULL;
968         int                      rc;
969
970         ENTRY;
971
972         /* Since the request might also come from lprocfs, so we need
973          * sync this with client_disconnect_export Bug15684 */
974         down_read(&exp->exp_obd->u.cli.cl_sem);
975         if (exp->exp_obd->u.cli.cl_import)
976                 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
977         up_read(&exp->exp_obd->u.cli.cl_sem);
978         if (!imp)
979                 RETURN(-ENODEV);
980
981         /* We could possibly pass max_age in the request (as an absolute
982          * timestamp or a "seconds.usec ago") so the target can avoid doing
983          * extra calls into the filesystem if that isn't necessary (e.g.
984          * during mount that would help a bit).  Having relative timestamps
985          * is not so great if request processing is slow, while absolute
986          * timestamps are not ideal because they need time synchronization. */
987         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
988
989         class_import_put(imp);
990
991         if (req == NULL)
992                 RETURN(-ENOMEM);
993
994         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
995         if (rc) {
996                 ptlrpc_request_free(req);
997                 RETURN(rc);
998         }
999         ptlrpc_request_set_replen(req);
1000         req->rq_request_portal = OST_CREATE_PORTAL;
1001         ptlrpc_at_set_req_timeout(req);
1002
1003         if (flags & OBD_STATFS_NODELAY) {
1004                 /* procfs requests not want stat in wait for avoid deadlock */
1005                 req->rq_no_resend = 1;
1006                 req->rq_no_delay = 1;
1007         }
1008
1009         rc = ptlrpc_queue_wait(req);
1010         if (rc)
1011                 GOTO(out, rc);
1012
1013         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1014         if (msfs == NULL)
1015                 GOTO(out, rc = -EPROTO);
1016
1017         *osfs = *msfs;
1018
1019         EXIT;
1020 out:
1021         ptlrpc_req_finished(req);
1022         return rc;
1023 }
1024
1025 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
1026                             enum obd_import_event event)
1027 {
1028         struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1029
1030         switch (event) {
1031         case IMP_EVENT_DISCON:
1032                 d->opd_got_disconnected = 1;
1033                 d->opd_imp_connected = 0;
1034                 if (d->opd_connect_mdt)
1035                         break;
1036
1037                 if (d->opd_pre != NULL) {
1038                         osp_pre_update_status(d, -ENODEV);
1039                         wake_up(&d->opd_pre_waitq);
1040                 }
1041
1042                 CDEBUG(D_HA, "got disconnected\n");
1043                 break;
1044         case IMP_EVENT_INACTIVE:
1045                 d->opd_imp_active = 0;
1046                 if (d->opd_connect_mdt)
1047                         break;
1048
1049                 if (d->opd_pre != NULL) {
1050                         osp_pre_update_status(d, -ENODEV);
1051                         wake_up(&d->opd_pre_waitq);
1052                 }
1053
1054                 CDEBUG(D_HA, "got inactive\n");
1055                 break;
1056         case IMP_EVENT_ACTIVE:
1057                 d->opd_imp_active = 1;
1058
1059                 if (osp_prepare_fid_client(d) != 0)
1060                         break;
1061
1062                 if (d->opd_got_disconnected)
1063                         d->opd_new_connection = 1;
1064                 d->opd_imp_connected = 1;
1065                 d->opd_imp_seen_connected = 1;
1066                 if (d->opd_connect_mdt)
1067                         break;
1068
1069                 if (d->opd_pre != NULL)
1070                         wake_up(&d->opd_pre_waitq);
1071
1072                 __osp_sync_check_for_work(d);
1073                 CDEBUG(D_HA, "got connected\n");
1074                 break;
1075         case IMP_EVENT_INVALIDATE:
1076                 if (obd->obd_namespace == NULL)
1077                         break;
1078                 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1079                 break;
1080         case IMP_EVENT_OCD:
1081         case IMP_EVENT_DEACTIVATE:
1082         case IMP_EVENT_ACTIVATE:
1083                 break;
1084         default:
1085                 CERROR("%s: unsupported import event: %#x\n",
1086                        obd->obd_name, event);
1087         }
1088         return 0;
1089 }
1090
1091 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1092                          void *karg, void *uarg)
1093 {
1094         struct obd_device       *obd = exp->exp_obd;
1095         struct osp_device       *d;
1096         struct obd_ioctl_data   *data = karg;
1097         int                      rc = 0;
1098
1099         ENTRY;
1100
1101         LASSERT(obd->obd_lu_dev);
1102         d = lu2osp_dev(obd->obd_lu_dev);
1103         LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1104
1105         if (!try_module_get(THIS_MODULE)) {
1106                 CERROR("%s: can't get module. Is it alive?", obd->obd_name);
1107                 return -EINVAL;
1108         }
1109
1110         switch (cmd) {
1111         case OBD_IOC_CLIENT_RECOVER:
1112                 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1113                                            data->ioc_inlbuf1, 0);
1114                 if (rc > 0)
1115                         rc = 0;
1116                 break;
1117         case IOC_OSC_SET_ACTIVE:
1118                 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1119                                               data->ioc_offset);
1120                 break;
1121         case OBD_IOC_PING_TARGET:
1122                 rc = ptlrpc_obd_ping(obd);
1123                 break;
1124         default:
1125                 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1126                        cmd, current_comm());
1127                 rc = -ENOTTY;
1128         }
1129         module_put(THIS_MODULE);
1130         return rc;
1131 }
1132
1133 static int osp_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1134                             __u32 keylen, void *key, __u32 *vallen, void *val,
1135                             struct lov_stripe_md *lsm)
1136 {
1137         int rc = -EINVAL;
1138
1139         if (KEY_IS(KEY_OSP_CONNECTED)) {
1140                 struct obd_device       *obd = exp->exp_obd;
1141                 struct osp_device       *osp;
1142
1143                 if (!obd->obd_set_up || obd->obd_stopping)
1144                         RETURN(-EAGAIN);
1145
1146                 osp = lu2osp_dev(obd->obd_lu_dev);
1147                 LASSERT(osp);
1148                 /*
1149                  * 1.8/2.0 behaviour is that OST being connected once at least
1150                  * is considered "healthy". and one "healthy" OST is enough to
1151                  * allow lustre clients to connect to MDS
1152                  */
1153                 RETURN(!osp->opd_imp_seen_connected);
1154         }
1155
1156         RETURN(rc);
1157 }
1158
1159 int osp_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
1160                   struct md_op_data *op_data)
1161 {
1162         struct client_obd *cli = &exp->exp_obd->u.cli;
1163         struct lu_client_seq *seq = cli->cl_seq;
1164
1165         ENTRY;
1166         RETURN(seq_client_alloc_fid(NULL, seq, fid));
1167 }
1168
1169 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1170 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1171 static void osp_key_exit(const struct lu_context *ctx,
1172                          struct lu_context_key *key, void *data)
1173 {
1174         struct osp_thread_info *info = data;
1175
1176         info->osi_attr.la_valid = 0;
1177 }
1178
1179 struct lu_context_key osp_thread_key = {
1180         .lct_tags = LCT_MD_THREAD,
1181         .lct_init = osp_key_init,
1182         .lct_fini = osp_key_fini,
1183         .lct_exit = osp_key_exit
1184 };
1185
1186 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1187 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1188
1189 struct lu_context_key osp_txn_key = {
1190         .lct_tags = LCT_OSP_THREAD | LCT_TX_HANDLE,
1191         .lct_init = osp_txn_key_init,
1192         .lct_fini = osp_txn_key_fini
1193 };
1194 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1195
1196 static struct lu_device_type_operations osp_device_type_ops = {
1197         .ldto_init           = osp_type_init,
1198         .ldto_fini           = osp_type_fini,
1199
1200         .ldto_start          = osp_type_start,
1201         .ldto_stop           = osp_type_stop,
1202
1203         .ldto_device_alloc   = osp_device_alloc,
1204         .ldto_device_free    = osp_device_free,
1205
1206         .ldto_device_fini    = osp_device_fini
1207 };
1208
1209 static struct lu_device_type osp_device_type = {
1210         .ldt_tags     = LU_DEVICE_DT,
1211         .ldt_name     = LUSTRE_OSP_NAME,
1212         .ldt_ops      = &osp_device_type_ops,
1213         .ldt_ctx_tags = LCT_MD_THREAD | LCT_DT_THREAD,
1214 };
1215
1216 static struct obd_ops osp_obd_device_ops = {
1217         .o_owner        = THIS_MODULE,
1218         .o_add_conn     = client_import_add_conn,
1219         .o_del_conn     = client_import_del_conn,
1220         .o_reconnect    = osp_reconnect,
1221         .o_connect      = osp_obd_connect,
1222         .o_disconnect   = osp_obd_disconnect,
1223         .o_get_info     = osp_obd_get_info,
1224         .o_import_event = osp_import_event,
1225         .o_iocontrol    = osp_iocontrol,
1226         .o_statfs       = osp_obd_statfs,
1227         .o_fid_init     = client_fid_init,
1228         .o_fid_fini     = client_fid_fini,
1229         .o_fid_alloc    = osp_fid_alloc,
1230 };
1231
1232 struct llog_operations osp_mds_ost_orig_logops;
1233
1234 static int __init osp_mod_init(void)
1235 {
1236         struct lprocfs_static_vars       lvars;
1237         cfs_proc_dir_entry_t            *osc_proc_dir;
1238         int                              rc;
1239
1240         rc = lu_kmem_init(osp_caches);
1241         if (rc)
1242                 return rc;
1243
1244         lprocfs_osp_init_vars(&lvars);
1245
1246         rc = class_register_type(&osp_obd_device_ops, NULL, NULL,
1247 #ifndef HAVE_ONLY_PROCFS_SEQ
1248                                 lvars.module_vars,
1249 #endif
1250                                 LUSTRE_OSP_NAME, &osp_device_type);
1251
1252         /* create "osc" entry in procfs for compatibility purposes */
1253         if (rc != 0) {
1254                 lu_kmem_fini(osp_caches);
1255                 return rc;
1256         }
1257
1258         lprocfs_lwp_init_vars(&lvars);
1259
1260         rc = class_register_type(&lwp_obd_device_ops, NULL, NULL,
1261 #ifndef HAVE_ONLY_PROCFS_SEQ
1262                                 lvars.module_vars,
1263 #endif
1264                                 LUSTRE_LWP_NAME, &lwp_device_type);
1265         if (rc != 0) {
1266                 class_unregister_type(LUSTRE_OSP_NAME);
1267                 lu_kmem_fini(osp_caches);
1268                 return rc;
1269         }
1270
1271         /* Note: add_rec/delcare_add_rec will be only used by catalogs */
1272         osp_mds_ost_orig_logops = llog_osd_ops;
1273         osp_mds_ost_orig_logops.lop_add = llog_cat_add_rec;
1274         osp_mds_ost_orig_logops.lop_declare_add = llog_cat_declare_add_rec;
1275
1276         osc_proc_dir = lprocfs_srch(proc_lustre_root, "osc");
1277         if (osc_proc_dir == NULL) {
1278                 osc_proc_dir = lprocfs_register("osc", proc_lustre_root, NULL,
1279                                                 NULL);
1280                 if (IS_ERR(osc_proc_dir))
1281                         CERROR("osp: can't create compat entry \"osc\": %d\n",
1282                                (int) PTR_ERR(osc_proc_dir));
1283         }
1284         return rc;
1285 }
1286
1287 static void __exit osp_mod_exit(void)
1288 {
1289         lprocfs_try_remove_proc_entry("osc", proc_lustre_root);
1290
1291         class_unregister_type(LUSTRE_LWP_NAME);
1292         class_unregister_type(LUSTRE_OSP_NAME);
1293         lu_kmem_fini(osp_caches);
1294 }
1295
1296 MODULE_AUTHOR("Intel, Inc. <http://www.intel.com/>");
1297 MODULE_DESCRIPTION("Lustre OST Proxy Device ("LUSTRE_OSP_NAME")");
1298 MODULE_LICENSE("GPL");
1299
1300 cfs_module(osp, LUSTRE_VERSION_STRING, osp_mod_init, osp_mod_exit);