Whamcloud - gitweb
e41133f3fe881e02574c9761cffa88918dc77b21
[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                 /* For data object, OSP obj would always be the top
81                  * object, i.e. hdr is always NULL, see lu_object_alloc.
82                  * But for metadata object, we always build the object
83                  * stack from MDT. i.e. mdt_object will be the top object
84                  * i.e.  hdr != NULL */
85                 if (hdr == NULL) {
86                         /* object for OST */
87                         h = &o->opo_header;
88                         lu_object_header_init(h);
89                         dt_object_init(&o->opo_obj, h, d);
90                         lu_object_add_top(h, l);
91                 } else {
92                         dt_object_init(&o->opo_obj, h, d);
93                 }
94
95                 l->lo_ops = &osp_lu_obj_ops;
96
97                 return l;
98         } else {
99                 return NULL;
100         }
101 }
102
103 static struct dt_object
104 *osp_find_or_create(const struct lu_env *env, struct osp_device *osp,
105                     struct lu_attr *attr, __u32 reg_id)
106 {
107         struct osp_thread_info *osi = osp_env_info(env);
108         struct dt_object_format dof = { 0 };
109         struct dt_object       *dto;
110         int                  rc;
111         ENTRY;
112
113         lu_local_obj_fid(&osi->osi_fid, reg_id);
114         attr->la_valid = LA_MODE;
115         attr->la_mode = S_IFREG | 0644;
116         dof.dof_type = DFT_REGULAR;
117         dto = dt_find_or_create(env, osp->opd_storage, &osi->osi_fid,
118                                 &dof, attr);
119         if (IS_ERR(dto))
120                 RETURN(dto);
121
122         rc = dt_attr_get(env, dto, attr, NULL);
123         if (rc) {
124                 CERROR("%s: can't be initialized: rc = %d\n",
125                        osp->opd_obd->obd_name, rc);
126                 lu_object_put(env, &dto->do_lu);
127                 RETURN(ERR_PTR(rc));
128         }
129         RETURN(dto);
130 }
131
132 static int osp_write_local_file(const struct lu_env *env,
133                                 struct osp_device *osp,
134                                 struct dt_object *dt_obj,
135                                 struct lu_buf *buf,
136                                 loff_t offset)
137 {
138         struct thandle *th;
139         int rc;
140
141         th = dt_trans_create(env, osp->opd_storage);
142         if (IS_ERR(th))
143                 RETURN(PTR_ERR(th));
144
145         rc = dt_declare_record_write(env, dt_obj, buf->lb_len, offset, th);
146         if (rc)
147                 GOTO(out, rc);
148         rc = dt_trans_start_local(env, osp->opd_storage, th);
149         if (rc)
150                 GOTO(out, rc);
151
152         rc = dt_record_write(env, dt_obj, buf, &offset, th);
153 out:
154         dt_trans_stop(env, osp->opd_storage, th);
155         RETURN(rc);
156 }
157
158 static int osp_init_last_objid(const struct lu_env *env, struct osp_device *osp)
159 {
160         struct osp_thread_info  *osi = osp_env_info(env);
161         struct lu_fid           *fid = &osp->opd_last_used_fid;
162         struct dt_object        *dto;
163         int                     rc;
164         ENTRY;
165
166         dto = osp_find_or_create(env, osp, &osi->osi_attr, MDD_LOV_OBJ_OID);
167         if (IS_ERR(dto))
168                 RETURN(PTR_ERR(dto));
169         /* object will be released in device cleanup path */
170         if (osi->osi_attr.la_size >=
171             sizeof(osi->osi_id) * (osp->opd_index + 1)) {
172                 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
173                                    osp->opd_index);
174                 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
175                 if (rc != 0)
176                         GOTO(out, rc);
177         } else {
178                 fid->f_oid = 0;
179                 osp_objid_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_oid,
180                                    osp->opd_index);
181                 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
182                                           osi->osi_off);
183         }
184         osp->opd_last_used_oid_file = dto;
185         RETURN(0);
186 out:
187         /* object will be released in device cleanup path */
188         CERROR("%s: can't initialize lov_objid: rc = %d\n",
189                osp->opd_obd->obd_name, rc);
190         lu_object_put(env, &dto->do_lu);
191         osp->opd_last_used_oid_file = NULL;
192         RETURN(rc);
193 }
194
195 static int osp_init_last_seq(const struct lu_env *env, struct osp_device *osp)
196 {
197         struct osp_thread_info  *osi = osp_env_info(env);
198         struct lu_fid           *fid = &osp->opd_last_used_fid;
199         struct dt_object        *dto;
200         int                     rc;
201         ENTRY;
202
203         dto = osp_find_or_create(env, osp, &osi->osi_attr, MDD_LOV_OBJ_OSEQ);
204         if (IS_ERR(dto))
205                 RETURN(PTR_ERR(dto));
206
207         /* object will be released in device cleanup path */
208         if (osi->osi_attr.la_size >=
209             sizeof(osi->osi_id) * (osp->opd_index + 1)) {
210                 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
211                                    osp->opd_index);
212                 rc = dt_record_read(env, dto, &osi->osi_lb, &osi->osi_off);
213                 if (rc != 0)
214                         GOTO(out, rc);
215         } else {
216                 fid->f_seq = 0;
217                 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off, &fid->f_seq,
218                                     osp->opd_index);
219                 rc = osp_write_local_file(env, osp, dto, &osi->osi_lb,
220                                           osi->osi_off);
221         }
222         osp->opd_last_used_seq_file = dto;
223         RETURN(0);
224 out:
225         /* object will be released in device cleanup path */
226         CERROR("%s: can't initialize lov_seq: rc = %d\n",
227                osp->opd_obd->obd_name, rc);
228         lu_object_put(env, &dto->do_lu);
229         osp->opd_last_used_seq_file = NULL;
230         RETURN(rc);
231 }
232
233 static int osp_last_used_init(const struct lu_env *env, struct osp_device *osp)
234 {
235         struct osp_thread_info *osi = osp_env_info(env);
236         int                  rc;
237         ENTRY;
238
239         fid_zero(&osp->opd_last_used_fid);
240         rc = osp_init_last_objid(env, osp);
241         if (rc < 0) {
242                 CERROR("%s: Can not get ids %d from old objid!\n",
243                        osp->opd_obd->obd_name, rc);
244                 RETURN(rc);
245         }
246
247         rc = osp_init_last_seq(env, osp);
248         if (rc < 0) {
249                 CERROR("%s: Can not get ids %d from old objid!\n",
250                        osp->opd_obd->obd_name, rc);
251                 GOTO(out, rc);
252         }
253
254         if (fid_oid(&osp->opd_last_used_fid) != 0 &&
255             fid_seq(&osp->opd_last_used_fid) == 0) {
256                 /* Just upgrade from the old version,
257                  * set the seq to be IDIF */
258                 osp->opd_last_used_fid.f_seq =
259                    fid_idif_seq(fid_oid(&osp->opd_last_used_fid),
260                                 osp->opd_index);
261                 osp_objseq_buf_prep(&osi->osi_lb, &osi->osi_off,
262                                     &osp->opd_last_used_fid.f_seq,
263                                     osp->opd_index);
264                 rc = osp_write_local_file(env, osp, osp->opd_last_used_seq_file,
265                                           &osi->osi_lb, osi->osi_off);
266                 if (rc) {
267                         CERROR("%s : Can not write seq file: rc = %d\n",
268                                osp->opd_obd->obd_name, rc);
269                         GOTO(out, rc);
270                 }
271         }
272
273         if (!fid_is_zero(&osp->opd_last_used_fid) &&
274                  !fid_is_sane(&osp->opd_last_used_fid)) {
275                 CERROR("%s: Got invalid FID "DFID"\n", osp->opd_obd->obd_name,
276                         PFID(&osp->opd_last_used_fid));
277                 GOTO(out, rc = -EINVAL);
278         }
279
280         CDEBUG(D_INFO, "%s: Init last used fid "DFID"\n",
281                osp->opd_obd->obd_name, PFID(&osp->opd_last_used_fid));
282 out:
283         if (rc != 0) {
284                 if (osp->opd_last_used_oid_file != NULL) {
285                         lu_object_put(env, &osp->opd_last_used_oid_file->do_lu);
286                         osp->opd_last_used_oid_file = NULL;
287                 }
288                 if (osp->opd_last_used_seq_file != NULL) {
289                         lu_object_put(env, &osp->opd_last_used_seq_file->do_lu);
290                         osp->opd_last_used_seq_file = NULL;
291                 }
292         }
293
294         RETURN(rc);
295 }
296
297 static void osp_last_used_fini(const struct lu_env *env, struct osp_device *d)
298 {
299         /* release last_used file */
300         if (d->opd_last_used_oid_file != NULL) {
301                 lu_object_put(env, &d->opd_last_used_oid_file->do_lu);
302                 d->opd_last_used_oid_file = NULL;
303         }
304
305         if (d->opd_last_used_seq_file != NULL) {
306                 lu_object_put(env, &d->opd_last_used_seq_file->do_lu);
307                 d->opd_last_used_seq_file = NULL;
308         }
309 }
310
311 static int osp_disconnect(struct osp_device *d)
312 {
313         struct obd_import *imp;
314         int rc = 0;
315
316         imp = d->opd_obd->u.cli.cl_import;
317
318         /* Mark import deactivated now, so we don't try to reconnect if any
319          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
320          * fully deactivate the import, or that would drop all requests. */
321         LASSERT(imp != NULL);
322         spin_lock(&imp->imp_lock);
323         imp->imp_deactive = 1;
324         spin_unlock(&imp->imp_lock);
325
326         ptlrpc_deactivate_import(imp);
327
328         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
329          * delete it regardless.  (It's safe to delete an import that was
330          * never added.) */
331         (void)ptlrpc_pinger_del_import(imp);
332
333         rc = ptlrpc_disconnect_import(imp, 0);
334         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
335                 rc = 0;
336         if (rc)
337                 CERROR("%s: can't disconnect: rc = %d\n",
338                        d->opd_obd->obd_name, rc);
339
340         ptlrpc_invalidate_import(imp);
341
342         RETURN(rc);
343 }
344
345 static int osp_shutdown(const struct lu_env *env, struct osp_device *d)
346 {
347         int                      rc = 0;
348         ENTRY;
349
350         LASSERT(env);
351
352         rc = osp_disconnect(d);
353
354         if (!d->opd_connect_mdt) {
355                 /* stop precreate thread */
356                 osp_precreate_fini(d);
357
358                 /* stop sync thread */
359                 osp_sync_fini(d);
360
361                 /* release last_used file */
362                 osp_last_used_fini(env, d);
363         }
364
365         obd_fid_fini(d->opd_obd);
366
367         RETURN(rc);
368 }
369
370 static int osp_process_config(const struct lu_env *env,
371                               struct lu_device *dev, struct lustre_cfg *lcfg)
372 {
373         struct osp_device               *d = lu2osp_dev(dev);
374         struct obd_device               *obd = d->opd_obd;
375         int                              rc;
376
377         ENTRY;
378
379         switch (lcfg->lcfg_command) {
380         case LCFG_PRE_CLEANUP:
381                 rc = osp_disconnect(d);
382                 break;
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                 LASSERT(obd);
389                 rc = class_process_proc_seq_param(PARAM_OSC, obd->obd_vars,
390                                                   lcfg, obd);
391                 if (rc > 0)
392                         rc = 0;
393                 if (rc == -ENOSYS) {
394                         /* class_process_proc_param() haven't found matching
395                          * parameter and returned ENOSYS so that layer(s)
396                          * below could use that. But OSP is the bottom, so
397                          * just ignore it */
398                         CERROR("%s: unknown param %s\n",
399                                (char *)lustre_cfg_string(lcfg, 0),
400                                (char *)lustre_cfg_string(lcfg, 1));
401                         rc = 0;
402                 }
403                 break;
404         default:
405                 CERROR("%s: unknown command %u\n",
406                        (char *)lustre_cfg_string(lcfg, 0), lcfg->lcfg_command);
407                 rc = 0;
408                 break;
409         }
410
411         RETURN(rc);
412 }
413
414 static int osp_recovery_complete(const struct lu_env *env,
415                                  struct lu_device *dev)
416 {
417         struct osp_device       *osp = lu2osp_dev(dev);
418         int                      rc = 0;
419
420         ENTRY;
421         osp->opd_recovery_completed = 1;
422
423         if (!osp->opd_connect_mdt && osp->opd_pre != NULL)
424                 wake_up(&osp->opd_pre_waitq);
425         RETURN(rc);
426 }
427
428 const struct lu_device_operations osp_lu_ops = {
429         .ldo_object_alloc       = osp_object_alloc,
430         .ldo_process_config     = osp_process_config,
431         .ldo_recovery_complete  = osp_recovery_complete,
432 };
433
434 /**
435  * provides with statfs from corresponded OST
436  *
437  */
438 static int osp_statfs(const struct lu_env *env, struct dt_device *dev,
439                       struct obd_statfs *sfs)
440 {
441         struct osp_device *d = dt2osp_dev(dev);
442
443         ENTRY;
444
445         if (unlikely(d->opd_imp_active == 0))
446                 RETURN(-ENOTCONN);
447
448         if (d->opd_pre == NULL)
449                 RETURN(0);
450
451         /* return recently updated data */
452         *sfs = d->opd_statfs;
453
454         /*
455          * layer above osp (usually lod) can use ffree to estimate
456          * how many objects are available for immediate creation
457          */
458         spin_lock(&d->opd_pre_lock);
459         LASSERTF(fid_seq(&d->opd_pre_last_created_fid) ==
460                  fid_seq(&d->opd_pre_used_fid),
461                  "last_created "DFID", next_fid "DFID"\n",
462                  PFID(&d->opd_pre_last_created_fid),
463                  PFID(&d->opd_pre_used_fid));
464         sfs->os_fprecreated = fid_oid(&d->opd_pre_last_created_fid) -
465                               fid_oid(&d->opd_pre_used_fid);
466         sfs->os_fprecreated -= d->opd_pre_reserved;
467         spin_unlock(&d->opd_pre_lock);
468
469         LASSERT(sfs->os_fprecreated <= OST_MAX_PRECREATE * 2);
470
471         CDEBUG(D_OTHER, "%s: "LPU64" blocks, "LPU64" free, "LPU64" avail, "
472                LPU64" files, "LPU64" free files\n", d->opd_obd->obd_name,
473                sfs->os_blocks, sfs->os_bfree, sfs->os_bavail,
474                sfs->os_files, sfs->os_ffree);
475         RETURN(0);
476 }
477
478 static int osp_sync(const struct lu_env *env, struct dt_device *dev)
479 {
480         ENTRY;
481
482         /*
483          * XXX: wake up sync thread, command it to start flushing asap?
484          */
485
486         RETURN(0);
487 }
488
489 const struct dt_device_operations osp_dt_ops = {
490         .dt_statfs      = osp_statfs,
491         .dt_sync        = osp_sync,
492         .dt_trans_start = osp_trans_start,
493         .dt_trans_stop  = osp_trans_stop,
494 };
495
496 static int osp_connect_to_osd(const struct lu_env *env, struct osp_device *m,
497                               const char *nextdev)
498 {
499         struct obd_connect_data *data = NULL;
500         struct obd_device       *obd;
501         int                      rc;
502
503         ENTRY;
504
505         LASSERT(m->opd_storage_exp == NULL);
506
507         OBD_ALLOC_PTR(data);
508         if (data == NULL)
509                 RETURN(-ENOMEM);
510
511         obd = class_name2obd(nextdev);
512         if (obd == NULL) {
513                 CERROR("%s: can't locate next device: %s\n",
514                        m->opd_obd->obd_name, nextdev);
515                 GOTO(out, rc = -ENOTCONN);
516         }
517
518         rc = obd_connect(env, &m->opd_storage_exp, obd, &obd->obd_uuid, data,
519                          NULL);
520         if (rc) {
521                 CERROR("%s: cannot connect to next dev %s: rc = %d\n",
522                        m->opd_obd->obd_name, nextdev, rc);
523                 GOTO(out, rc);
524         }
525
526         m->opd_dt_dev.dd_lu_dev.ld_site =
527                 m->opd_storage_exp->exp_obd->obd_lu_dev->ld_site;
528         LASSERT(m->opd_dt_dev.dd_lu_dev.ld_site);
529         m->opd_storage = lu2dt_dev(m->opd_storage_exp->exp_obd->obd_lu_dev);
530
531 out:
532         OBD_FREE_PTR(data);
533         RETURN(rc);
534 }
535
536 static int osp_init0(const struct lu_env *env, struct osp_device *m,
537                      struct lu_device_type *ldt, struct lustre_cfg *cfg)
538 {
539         struct obd_device       *obd;
540         struct obd_import       *imp;
541         class_uuid_t            uuid;
542         char                    *src, *tgt, *mdt, *osdname = NULL;
543         int                     rc;
544         long                    idx;
545
546         ENTRY;
547
548         obd = class_name2obd(lustre_cfg_string(cfg, 0));
549         if (obd == NULL) {
550                 CERROR("Cannot find obd with name %s\n",
551                        lustre_cfg_string(cfg, 0));
552                 RETURN(-ENODEV);
553         }
554         m->opd_obd = obd;
555
556         /* There is no record in the MDT configuration for the local disk
557          * device, so we have to extract this from elsewhere in the profile.
558          * The only information we get at setup is from the OSC records:
559          * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
560          * Note that 1.8 generated configs are missing the -MDTxxxx part.
561          * We need to reconstruct the name of the underlying OSD from this:
562          * {fsname}-{svname}-osd, for example "lustre-MDT0000-osd".  We
563          * also need to determine the OST index from this - will be used
564          * to calculate the offset in shared lov_objids file later */
565
566         src = lustre_cfg_string(cfg, 0);
567         if (src == NULL)
568                 RETURN(-EINVAL);
569
570         tgt = strrchr(src, '-');
571         if (tgt == NULL) {
572                 CERROR("%s: invalid target name %s\n",
573                        m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
574                 RETURN(-EINVAL);
575         }
576
577         if (strncmp(tgt, "-osc", 4) == 0) {
578                 /* Old OSC name fsname-OSTXXXX-osc */
579                 for (tgt--; tgt > src && *tgt != '-'; tgt--)
580                         ;
581                 if (tgt == src) {
582                         CERROR("%s: invalid target name %s\n",
583                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
584                         RETURN(-EINVAL);
585                 }
586
587                 if (strncmp(tgt, "-OST", 4) != 0) {
588                         CERROR("%s: invalid target name %s\n",
589                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
590                         RETURN(-EINVAL);
591                 }
592
593                 idx = simple_strtol(tgt + 4, &mdt, 16);
594                 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
595                         CERROR("%s: invalid OST index in '%s'\n",
596                                m->opd_obd->obd_name, src);
597                         RETURN(-EINVAL);
598                 }
599                 m->opd_index = idx;
600                 m->opd_group = 0;
601                 idx = tgt - src;
602         } else {
603                 /* New OSC name fsname-OSTXXXX-osc-MDTXXXX */
604                 if (strncmp(tgt, "-MDT", 4) != 0 &&
605                     strncmp(tgt, "-OST", 4) != 0) {
606                         CERROR("%s: invalid target name %s\n",
607                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
608                         RETURN(-EINVAL);
609                 }
610
611                 idx = simple_strtol(tgt + 4, &mdt, 16);
612                 if (*mdt != '\0' || idx > INT_MAX || idx < 0) {
613                         CERROR("%s: invalid OST index in '%s'\n",
614                                m->opd_obd->obd_name, src);
615                         RETURN(-EINVAL);
616                 }
617
618                 /* Get MDT index from the name and set it to opd_group,
619                  * which will be used by OSP to connect with OST */
620                 m->opd_group = idx;
621                 if (tgt - src <= 12) {
622                         CERROR("%s: invalid mdt index retrieve from %s\n",
623                                m->opd_obd->obd_name, lustre_cfg_string(cfg, 0));
624                         RETURN(-EINVAL);
625                 }
626
627                 if (strncmp(tgt - 12, "-MDT", 4) == 0)
628                         m->opd_connect_mdt = 1;
629
630                 idx = simple_strtol(tgt - 8, &mdt, 16);
631                 if (mdt[0] != '-' || idx > INT_MAX || idx < 0) {
632                         CERROR("%s: invalid OST index in '%s'\n",
633                                m->opd_obd->obd_name, src);
634                         RETURN(-EINVAL);
635                 }
636
637                 m->opd_index = idx;
638                 idx = tgt - src - 12;
639         }
640         /* check the fsname length, and after this everything else will fit */
641         if (idx > MTI_NAME_MAXLEN) {
642                 CERROR("%s: fsname too long in '%s'\n",
643                        m->opd_obd->obd_name, src);
644                 RETURN(-EINVAL);
645         }
646
647         OBD_ALLOC(osdname, MAX_OBD_NAME);
648         if (osdname == NULL)
649                 RETURN(-ENOMEM);
650
651         memcpy(osdname, src, idx); /* copy just the fsname part */
652         osdname[idx] = '\0';
653
654         mdt = strstr(mdt, "-MDT");
655         if (mdt == NULL) /* 1.8 configs don't have "-MDT0000" at the end */
656                 strcat(osdname, "-MDT0000");
657         else
658                 strcat(osdname, mdt);
659         strcat(osdname, "-osd");
660         CDEBUG(D_HA, "%s: connect to %s (%s)\n", obd->obd_name, osdname, src);
661
662         if (m->opd_connect_mdt) {
663                 struct client_obd *cli = &m->opd_obd->u.cli;
664
665                 OBD_ALLOC(cli->cl_rpc_lock, sizeof(*cli->cl_rpc_lock));
666                 if (!cli->cl_rpc_lock)
667                         GOTO(out_fini, rc = -ENOMEM);
668                 osp_init_rpc_lock(cli->cl_rpc_lock);
669         }
670
671         m->opd_dt_dev.dd_lu_dev.ld_ops = &osp_lu_ops;
672         m->opd_dt_dev.dd_ops = &osp_dt_ops;
673         obd->obd_lu_dev = &m->opd_dt_dev.dd_lu_dev;
674
675         rc = osp_connect_to_osd(env, m, osdname);
676         if (rc)
677                 GOTO(out_fini, rc);
678
679         rc = ptlrpcd_addref();
680         if (rc)
681                 GOTO(out_disconnect, rc);
682
683         rc = client_obd_setup(obd, cfg);
684         if (rc) {
685                 CERROR("%s: can't setup obd: %d\n", m->opd_obd->obd_name, rc);
686                 GOTO(out_ref, rc);
687         }
688
689         osp_lprocfs_init(m);
690
691         rc = obd_fid_init(m->opd_obd, NULL, m->opd_connect_mdt ?
692                           LUSTRE_SEQ_METADATA : LUSTRE_SEQ_DATA);
693         if (rc) {
694                 CERROR("%s: fid init error: rc = %d\n",
695                        m->opd_obd->obd_name, rc);
696                 GOTO(out_proc, rc);
697         }
698
699         if (!m->opd_connect_mdt) {
700                 /* Initialize last id from the storage - will be
701                  * used in orphan cleanup. */
702                 rc = osp_last_used_init(env, m);
703                 if (rc)
704                         GOTO(out_proc, rc);
705
706
707                 /* Initialize precreation thread, it handles new
708                  * connections as well. */
709                 rc = osp_init_precreate(m);
710                 if (rc)
711                         GOTO(out_last_used, rc);
712                 /*
713                  * Initialize synhronization mechanism taking
714                  * care of propogating changes to OST in near
715                  * transactional manner.
716                  */
717                 rc = osp_sync_init(env, m);
718                 if (rc)
719                         GOTO(out_precreat, rc);
720         }
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         if (!m->opd_connect_mdt)
739                 /* stop sync thread */
740                 osp_sync_fini(m);
741 out_precreat:
742         /* stop precreate thread */
743         if (!m->opd_connect_mdt)
744                 osp_precreate_fini(m);
745 out_last_used:
746         if (!m->opd_connect_mdt)
747                 osp_last_used_fini(env, m);
748 out_proc:
749         ptlrpc_lprocfs_unregister_obd(obd);
750         lprocfs_obd_cleanup(obd);
751         if (m->opd_symlink)
752                 lprocfs_remove(&m->opd_symlink);
753         client_obd_cleanup(obd);
754 out_ref:
755         ptlrpcd_decref();
756 out_disconnect:
757         if (m->opd_connect_mdt) {
758                 struct client_obd *cli = &m->opd_obd->u.cli;
759                 if (cli->cl_rpc_lock != NULL) {
760                         OBD_FREE_PTR(cli->cl_rpc_lock);
761                         cli->cl_rpc_lock = NULL;
762                 }
763         }
764         obd_disconnect(m->opd_storage_exp);
765 out_fini:
766         if (osdname)
767                 OBD_FREE(osdname, MAX_OBD_NAME);
768         RETURN(rc);
769 }
770
771 static struct lu_device *osp_device_free(const struct lu_env *env,
772                                          struct lu_device *lu)
773 {
774         struct osp_device *m = lu2osp_dev(lu);
775
776         ENTRY;
777
778         if (cfs_atomic_read(&lu->ld_ref) && lu->ld_site) {
779                 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
780                 lu_site_print(env, lu->ld_site, &msgdata, lu_cdebug_printer);
781         }
782         dt_device_fini(&m->opd_dt_dev);
783         OBD_FREE_PTR(m);
784         RETURN(NULL);
785 }
786
787 static struct lu_device *osp_device_alloc(const struct lu_env *env,
788                                           struct lu_device_type *t,
789                                           struct lustre_cfg *lcfg)
790 {
791         struct osp_device *m;
792         struct lu_device  *l;
793
794         OBD_ALLOC_PTR(m);
795         if (m == NULL) {
796                 l = ERR_PTR(-ENOMEM);
797         } else {
798                 int rc;
799
800                 l = osp2lu_dev(m);
801                 dt_device_init(&m->opd_dt_dev, t);
802                 rc = osp_init0(env, m, t, lcfg);
803                 if (rc != 0) {
804                         osp_device_free(env, l);
805                         l = ERR_PTR(rc);
806                 }
807         }
808         return l;
809 }
810
811 static struct lu_device *osp_device_fini(const struct lu_env *env,
812                                          struct lu_device *d)
813 {
814         struct osp_device *m = lu2osp_dev(d);
815         struct obd_import *imp;
816         int                rc;
817
818         ENTRY;
819
820         if (m->opd_storage_exp)
821                 obd_disconnect(m->opd_storage_exp);
822
823         imp = m->opd_obd->u.cli.cl_import;
824
825         if (imp->imp_rq_pool) {
826                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
827                 imp->imp_rq_pool = NULL;
828         }
829
830         if (m->opd_symlink)
831                 lprocfs_remove(&m->opd_symlink);
832
833         LASSERT(m->opd_obd);
834         ptlrpc_lprocfs_unregister_obd(m->opd_obd);
835         lprocfs_obd_cleanup(m->opd_obd);
836
837         if (m->opd_connect_mdt) {
838                 struct client_obd *cli = &m->opd_obd->u.cli;
839                 if (cli->cl_rpc_lock != NULL) {
840                         OBD_FREE_PTR(cli->cl_rpc_lock);
841                         cli->cl_rpc_lock = NULL;
842                 }
843         } else {
844                 if (m->opd_obd->obd_proc_private != NULL)
845                         lprocfs_remove((struct proc_dir_entry **)&m->opd_obd->obd_proc_private);
846         }
847
848         rc = client_obd_cleanup(m->opd_obd);
849         LASSERTF(rc == 0, "error %d\n", rc);
850
851         ptlrpcd_decref();
852
853         RETURN(NULL);
854 }
855
856 static int osp_reconnect(const struct lu_env *env,
857                          struct obd_export *exp, struct obd_device *obd,
858                          struct obd_uuid *cluuid,
859                          struct obd_connect_data *data,
860                          void *localdata)
861 {
862         return 0;
863 }
864
865 static int osp_prepare_fid_client(struct osp_device *osp)
866 {
867         int rc;
868
869         LASSERT(osp->opd_obd->u.cli.cl_seq != NULL);
870         if (osp->opd_obd->u.cli.cl_seq->lcs_exp != NULL)
871                 return 0;
872
873         LASSERT(osp->opd_exp != NULL);
874         osp->opd_obd->u.cli.cl_seq->lcs_exp =
875                                 class_export_get(osp->opd_exp);
876         if (osp->opd_pre == NULL)
877                 return 0;
878
879         /* Init fid for osp_precreate if necessary */
880         rc = osp_init_pre_fid(osp);
881         if (rc != 0) {
882                 class_export_put(osp->opd_exp);
883                 osp->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
884                 CERROR("%s: init pre fid error: rc = %d\n",
885                        osp->opd_obd->obd_name, rc);
886                 return rc;
887         }
888
889         return 0;
890 }
891
892 /*
893  * we use exports to track all LOD users
894  */
895 static int osp_obd_connect(const struct lu_env *env, struct obd_export **exp,
896                            struct obd_device *obd, struct obd_uuid *cluuid,
897                            struct obd_connect_data *data, void *localdata)
898 {
899         struct osp_device       *osp = lu2osp_dev(obd->obd_lu_dev);
900         struct obd_connect_data *ocd;
901         struct obd_import       *imp;
902         struct lustre_handle     conn;
903         int                      rc;
904
905         ENTRY;
906
907         CDEBUG(D_CONFIG, "connect #%d\n", osp->opd_connects);
908
909         rc = class_connect(&conn, obd, cluuid);
910         if (rc)
911                 RETURN(rc);
912
913         *exp = class_conn2export(&conn);
914         /* Why should there ever be more than 1 connect? */
915         osp->opd_connects++;
916         LASSERT(osp->opd_connects == 1);
917
918         osp->opd_exp = *exp;
919
920         imp = osp->opd_obd->u.cli.cl_import;
921         imp->imp_dlm_handle = conn;
922
923         LASSERT(data != NULL);
924         LASSERT(data->ocd_connect_flags & OBD_CONNECT_INDEX);
925         ocd = &imp->imp_connect_data;
926         *ocd = *data;
927
928         imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
929
930         ocd->ocd_version = LUSTRE_VERSION_CODE;
931         ocd->ocd_index = data->ocd_index;
932         imp->imp_connect_flags_orig = ocd->ocd_connect_flags;
933
934         rc = ptlrpc_connect_import(imp);
935         if (rc) {
936                 CERROR("%s: can't connect obd: rc = %d\n", obd->obd_name, rc);
937                 GOTO(out, rc);
938         }
939
940         ptlrpc_pinger_add_import(imp);
941 out:
942         RETURN(rc);
943 }
944
945 /*
946  * once last export (we don't count self-export) disappeared
947  * osp can be released
948  */
949 static int osp_obd_disconnect(struct obd_export *exp)
950 {
951         struct obd_device *obd = exp->exp_obd;
952         struct osp_device *osp = lu2osp_dev(obd->obd_lu_dev);
953         int                rc;
954         ENTRY;
955
956         /* Only disconnect the underlying layers on the final disconnect. */
957         LASSERT(osp->opd_connects == 1);
958         osp->opd_connects--;
959
960         rc = class_disconnect(exp);
961         if (rc) {
962                 CERROR("%s: class disconnect error: rc = %d\n",
963                        obd->obd_name, rc);
964                 RETURN(rc);
965         }
966
967         /* destroy the device */
968         class_manual_cleanup(obd);
969
970         RETURN(rc);
971 }
972
973 /*
974  * lprocfs helpers still use OBD API, let's keep obd_statfs() support
975  */
976 static int osp_obd_statfs(const struct lu_env *env, struct obd_export *exp,
977                           struct obd_statfs *osfs, __u64 max_age, __u32 flags)
978 {
979         struct obd_statfs       *msfs;
980         struct ptlrpc_request   *req;
981         struct obd_import       *imp = NULL;
982         int                      rc;
983
984         ENTRY;
985
986         /* Since the request might also come from lprocfs, so we need
987          * sync this with client_disconnect_export Bug15684 */
988         down_read(&exp->exp_obd->u.cli.cl_sem);
989         if (exp->exp_obd->u.cli.cl_import)
990                 imp = class_import_get(exp->exp_obd->u.cli.cl_import);
991         up_read(&exp->exp_obd->u.cli.cl_sem);
992         if (!imp)
993                 RETURN(-ENODEV);
994
995         /* We could possibly pass max_age in the request (as an absolute
996          * timestamp or a "seconds.usec ago") so the target can avoid doing
997          * extra calls into the filesystem if that isn't necessary (e.g.
998          * during mount that would help a bit).  Having relative timestamps
999          * is not so great if request processing is slow, while absolute
1000          * timestamps are not ideal because they need time synchronization. */
1001         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
1002
1003         class_import_put(imp);
1004
1005         if (req == NULL)
1006                 RETURN(-ENOMEM);
1007
1008         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
1009         if (rc) {
1010                 ptlrpc_request_free(req);
1011                 RETURN(rc);
1012         }
1013         ptlrpc_request_set_replen(req);
1014         req->rq_request_portal = OST_CREATE_PORTAL;
1015         ptlrpc_at_set_req_timeout(req);
1016
1017         if (flags & OBD_STATFS_NODELAY) {
1018                 /* procfs requests not want stat in wait for avoid deadlock */
1019                 req->rq_no_resend = 1;
1020                 req->rq_no_delay = 1;
1021         }
1022
1023         rc = ptlrpc_queue_wait(req);
1024         if (rc)
1025                 GOTO(out, rc);
1026
1027         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1028         if (msfs == NULL)
1029                 GOTO(out, rc = -EPROTO);
1030
1031         *osfs = *msfs;
1032
1033         EXIT;
1034 out:
1035         ptlrpc_req_finished(req);
1036         return rc;
1037 }
1038
1039 static int osp_import_event(struct obd_device *obd, struct obd_import *imp,
1040                             enum obd_import_event event)
1041 {
1042         struct osp_device *d = lu2osp_dev(obd->obd_lu_dev);
1043
1044         switch (event) {
1045         case IMP_EVENT_DISCON:
1046                 d->opd_got_disconnected = 1;
1047                 d->opd_imp_connected = 0;
1048                 if (d->opd_connect_mdt)
1049                         break;
1050
1051                 if (d->opd_pre != NULL) {
1052                         osp_pre_update_status(d, -ENODEV);
1053                         wake_up(&d->opd_pre_waitq);
1054                 }
1055
1056                 CDEBUG(D_HA, "got disconnected\n");
1057                 break;
1058         case IMP_EVENT_INACTIVE:
1059                 d->opd_imp_active = 0;
1060                 if (d->opd_connect_mdt)
1061                         break;
1062
1063                 if (d->opd_pre != NULL) {
1064                         osp_pre_update_status(d, -ENODEV);
1065                         wake_up(&d->opd_pre_waitq);
1066                 }
1067
1068                 CDEBUG(D_HA, "got inactive\n");
1069                 break;
1070         case IMP_EVENT_ACTIVE:
1071                 d->opd_imp_active = 1;
1072
1073                 if (osp_prepare_fid_client(d) != 0)
1074                         break;
1075
1076                 if (d->opd_got_disconnected)
1077                         d->opd_new_connection = 1;
1078                 d->opd_imp_connected = 1;
1079                 d->opd_imp_seen_connected = 1;
1080                 if (d->opd_connect_mdt)
1081                         break;
1082
1083                 if (d->opd_pre != NULL)
1084                         wake_up(&d->opd_pre_waitq);
1085
1086                 __osp_sync_check_for_work(d);
1087                 CDEBUG(D_HA, "got connected\n");
1088                 break;
1089         case IMP_EVENT_INVALIDATE:
1090                 if (obd->obd_namespace == NULL)
1091                         break;
1092                 ldlm_namespace_cleanup(obd->obd_namespace, LDLM_FL_LOCAL_ONLY);
1093                 break;
1094         case IMP_EVENT_OCD:
1095         case IMP_EVENT_DEACTIVATE:
1096         case IMP_EVENT_ACTIVATE:
1097                 break;
1098         default:
1099                 CERROR("%s: unsupported import event: %#x\n",
1100                        obd->obd_name, event);
1101         }
1102         return 0;
1103 }
1104
1105 static int osp_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1106                          void *karg, void *uarg)
1107 {
1108         struct obd_device       *obd = exp->exp_obd;
1109         struct osp_device       *d;
1110         struct obd_ioctl_data   *data = karg;
1111         int                      rc = 0;
1112
1113         ENTRY;
1114
1115         LASSERT(obd->obd_lu_dev);
1116         d = lu2osp_dev(obd->obd_lu_dev);
1117         LASSERT(d->opd_dt_dev.dd_ops == &osp_dt_ops);
1118
1119         if (!try_module_get(THIS_MODULE)) {
1120                 CERROR("%s: can't get module. Is it alive?", obd->obd_name);
1121                 return -EINVAL;
1122         }
1123
1124         switch (cmd) {
1125         case OBD_IOC_CLIENT_RECOVER:
1126                 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
1127                                            data->ioc_inlbuf1, 0);
1128                 if (rc > 0)
1129                         rc = 0;
1130                 break;
1131         case IOC_OSC_SET_ACTIVE:
1132                 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
1133                                               data->ioc_offset);
1134                 break;
1135         case OBD_IOC_PING_TARGET:
1136                 rc = ptlrpc_obd_ping(obd);
1137                 break;
1138         default:
1139                 CERROR("%s: unrecognized ioctl %#x by %s\n", obd->obd_name,
1140                        cmd, current_comm());
1141                 rc = -ENOTTY;
1142         }
1143         module_put(THIS_MODULE);
1144         return rc;
1145 }
1146
1147 static int osp_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1148                             __u32 keylen, void *key, __u32 *vallen, void *val,
1149                             struct lov_stripe_md *lsm)
1150 {
1151         int rc = -EINVAL;
1152
1153         if (KEY_IS(KEY_OSP_CONNECTED)) {
1154                 struct obd_device       *obd = exp->exp_obd;
1155                 struct osp_device       *osp;
1156
1157                 if (!obd->obd_set_up || obd->obd_stopping)
1158                         RETURN(-EAGAIN);
1159
1160                 osp = lu2osp_dev(obd->obd_lu_dev);
1161                 LASSERT(osp);
1162                 /*
1163                  * 1.8/2.0 behaviour is that OST being connected once at least
1164                  * is considered "healthy". and one "healthy" OST is enough to
1165                  * allow lustre clients to connect to MDS
1166                  */
1167                 RETURN(!osp->opd_imp_seen_connected);
1168         }
1169
1170         RETURN(rc);
1171 }
1172
1173 /* context key constructor/destructor: mdt_key_init, mdt_key_fini */
1174 LU_KEY_INIT_FINI(osp, struct osp_thread_info);
1175 static void osp_key_exit(const struct lu_context *ctx,
1176                          struct lu_context_key *key, void *data)
1177 {
1178         struct osp_thread_info *info = data;
1179
1180         info->osi_attr.la_valid = 0;
1181 }
1182
1183 struct lu_context_key osp_thread_key = {
1184         .lct_tags = LCT_MD_THREAD,
1185         .lct_init = osp_key_init,
1186         .lct_fini = osp_key_fini,
1187         .lct_exit = osp_key_exit
1188 };
1189
1190 /* context key constructor/destructor: mdt_txn_key_init, mdt_txn_key_fini */
1191 LU_KEY_INIT_FINI(osp_txn, struct osp_txn_info);
1192
1193 struct lu_context_key osp_txn_key = {
1194         .lct_tags = LCT_OSP_THREAD,
1195         .lct_init = osp_txn_key_init,
1196         .lct_fini = osp_txn_key_fini
1197 };
1198 LU_TYPE_INIT_FINI(osp, &osp_thread_key, &osp_txn_key);
1199
1200 static struct lu_device_type_operations osp_device_type_ops = {
1201         .ldto_init           = osp_type_init,
1202         .ldto_fini           = osp_type_fini,
1203
1204         .ldto_start          = osp_type_start,
1205         .ldto_stop           = osp_type_stop,
1206
1207         .ldto_device_alloc   = osp_device_alloc,
1208         .ldto_device_free    = osp_device_free,
1209
1210         .ldto_device_fini    = osp_device_fini
1211 };
1212
1213 static struct lu_device_type osp_device_type = {
1214         .ldt_tags     = LU_DEVICE_DT,
1215         .ldt_name     = LUSTRE_OSP_NAME,
1216         .ldt_ops      = &osp_device_type_ops,
1217         .ldt_ctx_tags = LCT_MD_THREAD
1218 };
1219
1220 static struct obd_ops osp_obd_device_ops = {
1221         .o_owner        = THIS_MODULE,
1222         .o_add_conn     = client_import_add_conn,
1223         .o_del_conn     = client_import_del_conn,
1224         .o_reconnect    = osp_reconnect,
1225         .o_connect      = osp_obd_connect,
1226         .o_disconnect   = osp_obd_disconnect,
1227         .o_get_info     = osp_obd_get_info,
1228         .o_import_event = osp_import_event,
1229         .o_iocontrol    = osp_iocontrol,
1230         .o_statfs       = osp_obd_statfs,
1231         .o_fid_init     = client_fid_init,
1232         .o_fid_fini     = client_fid_fini,
1233 };
1234
1235 struct llog_operations osp_mds_ost_orig_logops;
1236
1237 static int __init osp_mod_init(void)
1238 {
1239         int rc;
1240
1241         rc = lu_kmem_init(osp_caches);
1242         if (rc)
1243                 return rc;
1244
1245         rc = class_register_type(&osp_obd_device_ops, NULL, NULL,
1246 #ifndef HAVE_ONLY_PROCFS_SEQ
1247                                 NULL,
1248 #endif
1249                                 LUSTRE_OSP_NAME, &osp_device_type);
1250         if (rc != 0) {
1251                 lu_kmem_fini(osp_caches);
1252                 return rc;
1253         }
1254
1255         rc = class_register_type(&lwp_obd_device_ops, NULL, NULL,
1256 #ifndef HAVE_ONLY_PROCFS_SEQ
1257                                 NULL,
1258 #endif
1259                                 LUSTRE_LWP_NAME, &lwp_device_type);
1260         if (rc != 0) {
1261                 class_unregister_type(LUSTRE_OSP_NAME);
1262                 lu_kmem_fini(osp_caches);
1263                 return rc;
1264         }
1265
1266         /* Note: add_rec/delcare_add_rec will be only used by catalogs */
1267         osp_mds_ost_orig_logops = llog_osd_ops;
1268         osp_mds_ost_orig_logops.lop_add = llog_cat_add_rec;
1269         osp_mds_ost_orig_logops.lop_declare_add = llog_cat_declare_add_rec;
1270         return rc;
1271 }
1272
1273 static void __exit osp_mod_exit(void)
1274 {
1275         class_unregister_type(LUSTRE_LWP_NAME);
1276         class_unregister_type(LUSTRE_OSP_NAME);
1277         lu_kmem_fini(osp_caches);
1278 }
1279
1280 MODULE_AUTHOR("Intel, Inc. <http://www.intel.com/>");
1281 MODULE_DESCRIPTION("Lustre OST Proxy Device ("LUSTRE_OSP_NAME")");
1282 MODULE_LICENSE("GPL");
1283
1284 cfs_module(osp, LUSTRE_VERSION_STRING, osp_mod_init, osp_mod_exit);