Whamcloud - gitweb
LU-8901 misc: update Intel copyright messages for 2016
[fs/lustre-release.git] / lustre / lod / lod_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,
11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13  * GNU General Public License version 2 for more details.  A copy is
14  * included in the COPYING file that accompanied this code.
15  *
16  * You should have received a copy of the GNU General Public License
17  * along with this program; if not, write to the Free Software
18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright  2009 Sun Microsystems, Inc. All rights reserved
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2012, 2016, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/lod/lod_dev.c
33  *
34  * Lustre Logical Object Device
35  *
36  * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
37  * Author: Mikhail Pershin <mike.pershin@intel.com>
38  */
39 /**
40  * The Logical Object Device (LOD) layer manages access to striped
41  * objects (both regular files and directories). It implements the DT
42  * device and object APIs and is responsible for creating, storing,
43  * and loading striping information as an extended attribute of the
44  * underlying OSD object. LOD is the server side analog of the LOV and
45  * LMV layers on the client side.
46  *
47  * Metadata LU object stack (layers of the same compound LU object,
48  * all have the same FID):
49  *
50  *        MDT
51  *         |      MD API
52  *        MDD
53  *         |      DT API
54  *        LOD
55  *       /   \    DT API
56  *     OSD   OSP
57  *
58  * During LOD object initialization the localness or remoteness of the
59  * object FID dictates the choice between OSD and OSP.
60  *
61  * An LOD object (file or directory) with N stripes (each has a
62  * different FID):
63  *
64  *          LOD
65  *           |
66  *   +---+---+---+...+
67  *   |   |   |   |   |
68  *   S0  S1  S2  S3  S(N-1)  OS[DP] objects, seen as DT objects by LOD
69  *
70  * When upper layers must access an object's stripes (which are
71  * themselves OST or MDT LU objects) LOD finds these objects by their
72  * FIDs and stores them as an array of DT object pointers on the
73  * object. Declarations and operations on LOD objects are received by
74  * LOD (as DT object operations) and performed on the underlying
75  * OS[DP] object and (as needed) on the stripes. From the perspective
76  * of LOD, a stripe-less file (created by mknod() or open with
77  * O_LOV_DELAY_CREATE) is an object which does not yet have stripes,
78  * while a non-striped directory (created by mkdir()) is an object
79  * which will never have stripes.
80  *
81  * The LOD layer also implements a small subset of the OBD device API
82  * to support MDT stack initialization and finalization (an MDD device
83  * connects and disconnects itself to and from the underlying LOD
84  * device), and pool management. In turn LOD uses the OBD device API
85  * to connect it self to the underlying OSD, and to connect itself to
86  * OSP devices representing the MDTs and OSTs that bear the stripes of
87  * its objects.
88  */
89
90 #define DEBUG_SUBSYSTEM S_MDS
91
92 #include <linux/kthread.h>
93 #include <obd_class.h>
94 #include <md_object.h>
95 #include <lustre_fid.h>
96 #include <lustre_param.h>
97 #include <lustre_update.h>
98 #include <lustre_log.h>
99
100 #include "lod_internal.h"
101
102 static const char lod_update_log_name[] = "update_log";
103 static const char lod_update_log_dir_name[] = "update_log_dir";
104
105 /*
106  * Lookup target by FID.
107  *
108  * Lookup MDT/OST target index by FID. Type of the target can be
109  * specific or any.
110  *
111  * \param[in] env               LU environment provided by the caller
112  * \param[in] lod               lod device
113  * \param[in] fid               FID
114  * \param[out] tgt              result target index
115  * \param[in] type              expected type of the target:
116  *                              LU_SEQ_RANGE_{MDT,OST,ANY}
117  *
118  * \retval 0                    on success
119  * \retval negative             negated errno on error
120  **/
121 int lod_fld_lookup(const struct lu_env *env, struct lod_device *lod,
122                    const struct lu_fid *fid, __u32 *tgt, int *type)
123 {
124         struct lu_seq_range     range = { 0 };
125         struct lu_server_fld    *server_fld;
126         int rc;
127         ENTRY;
128
129         if (!fid_is_sane(fid)) {
130                 CERROR("%s: invalid FID "DFID"\n", lod2obd(lod)->obd_name,
131                        PFID(fid));
132                 RETURN(-EIO);
133         }
134
135         if (fid_is_idif(fid)) {
136                 *tgt = fid_idif_ost_idx(fid);
137                 *type = LU_SEQ_RANGE_OST;
138                 RETURN(0);
139         }
140
141         if (fid_is_update_log(fid) || fid_is_update_log_dir(fid)) {
142                 *tgt = fid_oid(fid);
143                 *type = LU_SEQ_RANGE_MDT;
144                 RETURN(0);
145         }
146
147         if (!lod->lod_initialized || (!fid_seq_in_fldb(fid_seq(fid)))) {
148                 LASSERT(lu_site2seq(lod2lu_dev(lod)->ld_site) != NULL);
149
150                 *tgt = lu_site2seq(lod2lu_dev(lod)->ld_site)->ss_node_id;
151                 *type = LU_SEQ_RANGE_MDT;
152                 RETURN(0);
153         }
154
155         server_fld = lu_site2seq(lod2lu_dev(lod)->ld_site)->ss_server_fld;
156         if (server_fld == NULL)
157                 RETURN(-EIO);
158
159         fld_range_set_type(&range, *type);
160         rc = fld_server_lookup(env, server_fld, fid_seq(fid), &range);
161         if (rc != 0)
162                 RETURN(rc);
163
164         *tgt = range.lsr_index;
165         *type = range.lsr_flags;
166
167         CDEBUG(D_INFO, "%s: got tgt %x for sequence: %#llx\n",
168                lod2obd(lod)->obd_name, *tgt, fid_seq(fid));
169
170         RETURN(0);
171 }
172
173 /* Slab for OSD object allocation */
174 struct kmem_cache *lod_object_kmem;
175
176 /* Slab for dt_txn_callback */
177 struct kmem_cache *lod_txn_callback_kmem;
178 static struct lu_kmem_descr lod_caches[] = {
179         {
180                 .ckd_cache = &lod_object_kmem,
181                 .ckd_name  = "lod_obj",
182                 .ckd_size  = sizeof(struct lod_object)
183         },
184         {
185                 .ckd_cache = &lod_txn_callback_kmem,
186                 .ckd_name  = "lod_txn_callback",
187                 .ckd_size  = sizeof(struct dt_txn_callback)
188         },
189         {
190                 .ckd_cache = NULL
191         }
192 };
193
194 static struct lu_device *lod_device_fini(const struct lu_env *env,
195                                          struct lu_device *d);
196
197 /**
198  * Implementation of lu_device_operations::ldo_object_alloc() for LOD
199  *
200  * Allocates and initializes LOD's slice in the given object.
201  *
202  * see include/lu_object.h for the details.
203  */
204 static struct lu_object *lod_object_alloc(const struct lu_env *env,
205                                           const struct lu_object_header *hdr,
206                                           struct lu_device *dev)
207 {
208         struct lod_object       *lod_obj;
209         struct lu_object        *lu_obj;
210         ENTRY;
211
212         OBD_SLAB_ALLOC_PTR_GFP(lod_obj, lod_object_kmem, GFP_NOFS);
213         if (lod_obj == NULL)
214                 RETURN(ERR_PTR(-ENOMEM));
215
216         lu_obj = lod2lu_obj(lod_obj);
217         dt_object_init(&lod_obj->ldo_obj, NULL, dev);
218         lod_obj->ldo_obj.do_ops = &lod_obj_ops;
219         lu_obj->lo_ops = &lod_lu_obj_ops;
220
221         RETURN(lu_obj);
222 }
223
224 /**
225  * Process the config log for all sub device.
226  *
227  * The function goes through all the targets in the given table
228  * and apply given configuration command on to the targets.
229  * Used to cleanup the targets at unmount.
230  *
231  * \param[in] env               LU environment provided by the caller
232  * \param[in] lod               lod device
233  * \param[in] ltd               target's table to go through
234  * \param[in] lcfg              configuration command to apply
235  *
236  * \retval 0                    on success
237  * \retval negative             negated errno on error
238  **/
239 static int lod_sub_process_config(const struct lu_env *env,
240                                  struct lod_device *lod,
241                                  struct lod_tgt_descs *ltd,
242                                  struct lustre_cfg *lcfg)
243 {
244         struct lu_device  *next;
245         int rc = 0;
246         unsigned int i;
247
248         lod_getref(ltd);
249         if (ltd->ltd_tgts_size <= 0) {
250                 lod_putref(lod, ltd);
251                 return 0;
252         }
253         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
254                 struct lod_tgt_desc *tgt;
255                 int rc1;
256
257                 tgt = LTD_TGT(ltd, i);
258                 LASSERT(tgt && tgt->ltd_tgt);
259                 next = &tgt->ltd_tgt->dd_lu_dev;
260                 rc1 = next->ld_ops->ldo_process_config(env, next, lcfg);
261                 if (rc1) {
262                         CERROR("%s: error cleaning up LOD index %u: cmd %#x"
263                                ": rc = %d\n", lod2obd(lod)->obd_name, i,
264                                lcfg->lcfg_command, rc1);
265                         rc = rc1;
266                 }
267         }
268         lod_putref(lod, ltd);
269         return rc;
270 }
271
272 struct lod_recovery_data {
273         struct lod_device       *lrd_lod;
274         struct lod_tgt_desc     *lrd_ltd;
275         struct ptlrpc_thread    *lrd_thread;
276         __u32                   lrd_idx;
277 };
278
279
280 /**
281  * process update recovery record
282  *
283  * Add the update recovery recode to the update recovery list in
284  * lod_recovery_data. Then the recovery thread (target_recovery_thread)
285  * will redo these updates.
286  *
287  * \param[in]env        execution environment
288  * \param[in]llh        log handle of update record
289  * \param[in]rec        update record to be replayed
290  * \param[in]data       update recovery data which holds the necessary
291  *                      arguments for recovery (see struct lod_recovery_data)
292  *
293  * \retval              0 if the record is processed successfully.
294  * \retval              negative errno if the record processing fails.
295  */
296 static int lod_process_recovery_updates(const struct lu_env *env,
297                                         struct llog_handle *llh,
298                                         struct llog_rec_hdr *rec,
299                                         void *data)
300 {
301         struct lod_recovery_data        *lrd = data;
302         struct llog_cookie      *cookie = &lod_env_info(env)->lti_cookie;
303         struct lu_target                *lut;
304         __u32                           index = 0;
305         ENTRY;
306
307         if (lrd->lrd_ltd == NULL) {
308                 int rc;
309
310                 rc = lodname2mdt_index(lod2obd(lrd->lrd_lod)->obd_name, &index);
311                 if (rc != 0)
312                         return rc;
313         } else {
314                 index = lrd->lrd_ltd->ltd_index;
315         }
316
317         if (rec->lrh_len !=
318                 llog_update_record_size((struct llog_update_record *)rec)) {
319                 CERROR("%s broken update record! index %u "DOSTID":%u :"
320                        " rc = %d\n", lod2obd(lrd->lrd_lod)->obd_name, index,
321                        POSTID(&llh->lgh_id.lgl_oi), rec->lrh_index, -EIO);
322                 return -EINVAL;
323         }
324
325         cookie->lgc_lgl = llh->lgh_id;
326         cookie->lgc_index = rec->lrh_index;
327         cookie->lgc_subsys = LLOG_UPDATELOG_ORIG_CTXT;
328
329         CDEBUG(D_HA, "%s: process recovery updates "DOSTID":%u\n",
330                lod2obd(lrd->lrd_lod)->obd_name,
331                POSTID(&llh->lgh_id.lgl_oi), rec->lrh_index);
332         lut = lod2lu_dev(lrd->lrd_lod)->ld_site->ls_tgt;
333
334         if (lut->lut_obd->obd_stopping ||
335             lut->lut_obd->obd_abort_recovery)
336                 return -ESHUTDOWN;
337
338         return insert_update_records_to_replay_list(lut->lut_tdtd,
339                                         (struct llog_update_record *)rec,
340                                         cookie, index);
341 }
342
343 /**
344  * recovery thread for update log
345  *
346  * Start recovery thread and prepare the sub llog, then it will retrieve
347  * the update records from the correpondent MDT and do recovery.
348  *
349  * \param[in] arg       pointer to the recovery data
350  *
351  * \retval              0 if recovery succeeds
352  * \retval              negative errno if recovery failed.
353  */
354 static int lod_sub_recovery_thread(void *arg)
355 {
356         struct lod_recovery_data        *lrd = arg;
357         struct lod_device               *lod = lrd->lrd_lod;
358         struct dt_device                *dt;
359         struct ptlrpc_thread            *thread = lrd->lrd_thread;
360         struct llog_ctxt                *ctxt = NULL;
361         struct lu_env                   env;
362         struct lu_target *lut;
363
364
365         int                             rc;
366         ENTRY;
367
368         thread->t_flags = SVC_RUNNING;
369         wake_up(&thread->t_ctl_waitq);
370
371         rc = lu_env_init(&env, LCT_LOCAL | LCT_MD_THREAD);
372         if (rc != 0) {
373                 OBD_FREE_PTR(lrd);
374                 CERROR("%s: can't initialize env: rc = %d\n",
375                        lod2obd(lod)->obd_name, rc);
376                 RETURN(rc);
377         }
378
379         lut = lod2lu_dev(lod)->ld_site->ls_tgt;
380         atomic_inc(&lut->lut_tdtd->tdtd_recovery_threads_count);
381         if (lrd->lrd_ltd == NULL)
382                 dt = lod->lod_child;
383         else
384                 dt = lrd->lrd_ltd->ltd_tgt;
385
386 again:
387         rc = lod_sub_prep_llog(&env, lod, dt, lrd->lrd_idx);
388         if (rc == 0) {
389                 /* Process the recovery record */
390                 ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
391                                         LLOG_UPDATELOG_ORIG_CTXT);
392                 LASSERT(ctxt != NULL);
393                 LASSERT(ctxt->loc_handle != NULL);
394
395                 rc = llog_cat_process(&env, ctxt->loc_handle,
396                                       lod_process_recovery_updates, lrd, 0, 0);
397         }
398
399         if (rc < 0) {
400                 struct lu_device *top_device;
401
402                 top_device = lod->lod_dt_dev.dd_lu_dev.ld_site->ls_top_dev;
403                 /* Because the remote target might failover at the same time,
404                  * let's retry here */
405                 if ((rc == -ETIMEDOUT || rc == -EAGAIN || rc == -EIO) &&
406                      dt != lod->lod_child &&
407                     !top_device->ld_obd->obd_abort_recovery &&
408                     !top_device->ld_obd->obd_stopping) {
409                         if (ctxt != NULL) {
410                                 if (ctxt->loc_handle != NULL)
411                                         llog_cat_close(&env,
412                                                        ctxt->loc_handle);
413                                 llog_ctxt_put(ctxt);
414                         }
415                         goto again;
416                 }
417
418                 CERROR("%s getting update log failed: rc = %d\n",
419                        dt->dd_lu_dev.ld_obd->obd_name, rc);
420                 llog_ctxt_put(ctxt);
421
422                 spin_lock(&top_device->ld_obd->obd_dev_lock);
423                 if (!top_device->ld_obd->obd_abort_recovery &&
424                     !top_device->ld_obd->obd_stopping)
425                         top_device->ld_obd->obd_abort_recovery = 1;
426                 spin_unlock(&top_device->ld_obd->obd_dev_lock);
427
428                 GOTO(out, rc);
429         }
430         llog_ctxt_put(ctxt);
431
432         CDEBUG(D_HA, "%s retrieve update log: rc = %d\n",
433                dt->dd_lu_dev.ld_obd->obd_name, rc);
434
435         if (lrd->lrd_ltd == NULL)
436                 lod->lod_child_got_update_log = 1;
437         else
438                 lrd->lrd_ltd->ltd_got_update_log = 1;
439
440         if (lod->lod_child_got_update_log) {
441                 struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
442                 struct lod_tgt_desc     *tgt = NULL;
443                 bool                    all_got_log = true;
444                 int                     i;
445
446                 cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
447                         tgt = LTD_TGT(ltd, i);
448                         if (!tgt->ltd_got_update_log) {
449                                 all_got_log = false;
450                                 break;
451                         }
452                 }
453
454                 if (all_got_log) {
455                         CDEBUG(D_HA, "%s got update logs from all MDTs.\n",
456                                lut->lut_obd->obd_name);
457                         lut->lut_tdtd->tdtd_replay_ready = 1;
458                         wake_up(&lut->lut_obd->obd_next_transno_waitq);
459                 }
460         }
461
462 out:
463         OBD_FREE_PTR(lrd);
464         thread->t_flags = SVC_STOPPED;
465         atomic_dec(&lut->lut_tdtd->tdtd_recovery_threads_count);
466         wake_up(&lut->lut_tdtd->tdtd_recovery_threads_waitq);
467         wake_up(&thread->t_ctl_waitq);
468         lu_env_fini(&env);
469         RETURN(rc);
470 }
471
472 /**
473  * finish sub llog context
474  *
475  * Stop update recovery thread for the sub device, then cleanup the
476  * correspondent llog ctxt.
477  *
478  * \param[in] env      execution environment
479  * \param[in] lod      lod device to do update recovery
480  * \param[in] thread   recovery thread on this sub device
481  */
482 void lod_sub_fini_llog(const struct lu_env *env,
483                        struct dt_device *dt, struct ptlrpc_thread *thread)
484 {
485         struct obd_device       *obd;
486         struct llog_ctxt        *ctxt;
487         ENTRY;
488
489         obd = dt->dd_lu_dev.ld_obd;
490         CDEBUG(D_INFO, "%s: finish sub llog\n", obd->obd_name);
491         /* Stop recovery thread first */
492         if (thread != NULL && thread->t_flags & SVC_RUNNING) {
493                 thread->t_flags = SVC_STOPPING;
494                 wake_up(&thread->t_ctl_waitq);
495                 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
496         }
497
498         ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
499         if (ctxt == NULL)
500                 RETURN_EXIT;
501
502         if (ctxt->loc_handle != NULL)
503                 llog_cat_close(env, ctxt->loc_handle);
504
505         llog_cleanup(env, ctxt);
506
507         RETURN_EXIT;
508 }
509
510 /**
511  * Extract MDT target index from a device name.
512  *
513  * a helper function to extract index from the given device name
514  * like "fsname-MDTxxxx-mdtlov"
515  *
516  * \param[in] lodname           device name
517  * \param[out] mdt_index        extracted index
518  *
519  * \retval 0            on success
520  * \retval -EINVAL      if the name is invalid
521  */
522 int lodname2mdt_index(char *lodname, __u32 *mdt_index)
523 {
524         unsigned long index;
525         char *ptr, *tmp;
526
527         /* 1.8 configs don't have "-MDT0000" at the end */
528         ptr = strstr(lodname, "-MDT");
529         if (ptr == NULL) {
530                 *mdt_index = 0;
531                 return 0;
532         }
533
534         ptr = strrchr(lodname, '-');
535         if (ptr == NULL) {
536                 CERROR("invalid MDT index in '%s'\n", lodname);
537                 return -EINVAL;
538         }
539
540         if (strncmp(ptr, "-mdtlov", 7) != 0) {
541                 CERROR("invalid MDT index in '%s'\n", lodname);
542                 return -EINVAL;
543         }
544
545         if ((unsigned long)ptr - (unsigned long)lodname <= 8) {
546                 CERROR("invalid MDT index in '%s'\n", lodname);
547                 return -EINVAL;
548         }
549
550         if (strncmp(ptr - 8, "-MDT", 4) != 0) {
551                 CERROR("invalid MDT index in '%s'\n", lodname);
552                 return -EINVAL;
553         }
554
555         index = simple_strtol(ptr - 4, &tmp, 16);
556         if (*tmp != '-' || index > INT_MAX) {
557                 CERROR("invalid MDT index in '%s'\n", lodname);
558                 return -EINVAL;
559         }
560         *mdt_index = index;
561         return 0;
562 }
563
564 /**
565  * Init sub llog context
566  *
567  * Setup update llog ctxt for update recovery threads, then start the
568  * recovery thread (lod_sub_recovery_thread) to read update llog from
569  * the correspondent MDT to do update recovery.
570  *
571  * \param[in] env       execution environment
572  * \param[in] lod       lod device to do update recovery
573  * \param[in] dt        sub dt device for which the recovery thread is
574  *
575  * \retval              0 if initialization succeeds.
576  * \retval              negative errno if initialization fails.
577  */
578 int lod_sub_init_llog(const struct lu_env *env, struct lod_device *lod,
579                       struct dt_device *dt)
580 {
581         struct obd_device               *obd;
582         struct lod_recovery_data        *lrd = NULL;
583         struct ptlrpc_thread            *thread;
584         struct task_struct              *task;
585         struct l_wait_info              lwi = { 0 };
586         struct lod_tgt_desc             *sub_ltd = NULL;
587         __u32                           index;
588         __u32                           master_index;
589         int                             rc;
590         ENTRY;
591
592         rc = lodname2mdt_index(lod2obd(lod)->obd_name, &master_index);
593         if (rc != 0)
594                 RETURN(rc);
595
596         OBD_ALLOC_PTR(lrd);
597         if (lrd == NULL)
598                 RETURN(-ENOMEM);
599
600         if (lod->lod_child == dt) {
601                 thread = &lod->lod_child_recovery_thread;
602                 index = master_index;
603         } else {
604                 struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
605                 struct lod_tgt_desc     *tgt = NULL;
606                 unsigned int            i;
607
608                 cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
609                         tgt = LTD_TGT(ltd, i);
610                         if (tgt->ltd_tgt == dt) {
611                                 index = tgt->ltd_index;
612                                 sub_ltd = tgt;
613                                 break;
614                         }
615                 }
616                 LASSERT(sub_ltd != NULL);
617                 OBD_ALLOC_PTR(sub_ltd->ltd_recovery_thread);
618                 if (sub_ltd->ltd_recovery_thread == NULL)
619                         GOTO(free_lrd, rc = -ENOMEM);
620
621                 thread = sub_ltd->ltd_recovery_thread;
622         }
623
624         CDEBUG(D_INFO, "%s init sub log %s\n", lod2obd(lod)->obd_name,
625                dt->dd_lu_dev.ld_obd->obd_name);
626         lrd->lrd_lod = lod;
627         lrd->lrd_ltd = sub_ltd;
628         lrd->lrd_thread = thread;
629         lrd->lrd_idx = index;
630         init_waitqueue_head(&thread->t_ctl_waitq);
631
632         obd = dt->dd_lu_dev.ld_obd;
633         obd->obd_lvfs_ctxt.dt = dt;
634         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_UPDATELOG_ORIG_CTXT,
635                         NULL, &llog_common_cat_ops);
636         if (rc < 0) {
637                 CERROR("%s: cannot setup updatelog llog: rc = %d\n",
638                        obd->obd_name, rc);
639                 GOTO(free_thread, rc);
640         }
641
642         /* Start the recovery thread */
643         task = kthread_run(lod_sub_recovery_thread, lrd, "lod%04x_rec%04x",
644                            master_index, index);
645         if (IS_ERR(task)) {
646                 rc = PTR_ERR(task);
647                 CERROR("%s: cannot start recovery thread: rc = %d\n",
648                        obd->obd_name, rc);
649                 GOTO(out_llog, rc);
650         }
651
652         l_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING ||
653                                           thread->t_flags & SVC_STOPPED, &lwi);
654
655         RETURN(0);
656 out_llog:
657         lod_sub_fini_llog(env, dt, thread);
658 free_thread:
659         if (lod->lod_child != dt) {
660                 OBD_FREE_PTR(sub_ltd->ltd_recovery_thread);
661                 sub_ltd->ltd_recovery_thread = NULL;
662         }
663 free_lrd:
664         OBD_FREE_PTR(lrd);
665         RETURN(rc);
666 }
667
668 /**
669  * Stop sub recovery thread
670  *
671  * Stop sub recovery thread on all subs.
672  *
673  * \param[in] env       execution environment
674  * \param[in] lod       lod device to do update recovery
675  */
676 static void lod_sub_stop_recovery_threads(const struct lu_env *env,
677                                           struct lod_device *lod)
678 {
679         struct lod_tgt_descs *ltd = &lod->lod_mdt_descs;
680         struct ptlrpc_thread    *thread;
681         unsigned int i;
682
683         /* Stop the update log commit cancel threads and finish master
684          * llog ctxt */
685         thread = &lod->lod_child_recovery_thread;
686         /* Stop recovery thread first */
687         if (thread != NULL && thread->t_flags & SVC_RUNNING) {
688                 thread->t_flags = SVC_STOPPING;
689                 wake_up(&thread->t_ctl_waitq);
690                 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
691         }
692
693         lod_getref(ltd);
694         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
695                 struct lod_tgt_desc     *tgt;
696
697                 tgt = LTD_TGT(ltd, i);
698                 thread = tgt->ltd_recovery_thread;
699                 if (thread != NULL && thread->t_flags & SVC_RUNNING) {
700                         thread->t_flags = SVC_STOPPING;
701                         wake_up(&thread->t_ctl_waitq);
702                         wait_event(thread->t_ctl_waitq,
703                                    thread->t_flags & SVC_STOPPED);
704                         OBD_FREE_PTR(tgt->ltd_recovery_thread);
705                         tgt->ltd_recovery_thread = NULL;
706                 }
707         }
708
709         lod_putref(lod, ltd);
710 }
711
712 /**
713  * finish all sub llog
714  *
715  * cleanup all of sub llog ctxt on the LOD.
716  *
717  * \param[in] env       execution environment
718  * \param[in] lod       lod device to do update recovery
719  */
720 static void lod_sub_fini_all_llogs(const struct lu_env *env,
721                                    struct lod_device *lod)
722 {
723         struct lod_tgt_descs *ltd = &lod->lod_mdt_descs;
724         unsigned int i;
725
726         /* Stop the update log commit cancel threads and finish master
727          * llog ctxt */
728         lod_sub_fini_llog(env, lod->lod_child,
729                           &lod->lod_child_recovery_thread);
730         lod_getref(ltd);
731         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
732                 struct lod_tgt_desc     *tgt;
733
734                 tgt = LTD_TGT(ltd, i);
735                 lod_sub_fini_llog(env, tgt->ltd_tgt,
736                                   tgt->ltd_recovery_thread);
737         }
738
739         lod_putref(lod, ltd);
740 }
741
742 static char *lod_show_update_logs_retrievers(void *data, int *size, int *count)
743 {
744         struct lod_device       *lod = (struct lod_device *)data;
745         struct lu_target        *lut = lod2lu_dev(lod)->ld_site->ls_tgt;
746         struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
747         struct lod_tgt_desc     *tgt = NULL;
748         char                    *buf;
749         int                      len = 0;
750         int                      rc;
751         int                      i;
752
753         *count = atomic_read(&lut->lut_tdtd->tdtd_recovery_threads_count);
754         if (*count == 0) {
755                 *size = 0;
756                 return NULL;
757         }
758
759         *size = 5 * *count + 1;
760         OBD_ALLOC(buf, *size);
761         if (buf == NULL)
762                 return NULL;
763
764         *count = 0;
765         memset(buf, 0, *size);
766
767         if (!lod->lod_child_got_update_log) {
768                 rc = lodname2mdt_index(lod2obd(lod)->obd_name, &i);
769                 LASSERTF(rc == 0, "Fail to parse target index: rc = %d\n", rc);
770
771                 rc = snprintf(buf + len, *size - len, " %04x", i);
772                 LASSERT(rc > 0);
773
774                 len += rc;
775                 (*count)++;
776         }
777
778         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
779                 tgt = LTD_TGT(ltd, i);
780                 if (!tgt->ltd_got_update_log) {
781                         rc = snprintf(buf + len, *size - len, " %04x", i);
782                         if (unlikely(rc <= 0))
783                                 break;
784
785                         len += rc;
786                         (*count)++;
787                 }
788         }
789
790         return buf;
791 }
792
793 /**
794  * Prepare distribute txn
795  *
796  * Prepare distribute txn structure for LOD
797  *
798  * \param[in] env       execution environment
799  * \param[in] lod_device  LOD device
800  *
801  * \retval              0 if preparation succeeds.
802  * \retval              negative errno if preparation fails.
803  */
804 static int lod_prepare_distribute_txn(const struct lu_env *env,
805                                       struct lod_device *lod)
806 {
807         struct target_distribute_txn_data *tdtd;
808         struct lu_target                  *lut;
809         int                               rc;
810         ENTRY;
811
812         /* Init update recovery data */
813         OBD_ALLOC_PTR(tdtd);
814         if (tdtd == NULL)
815                 RETURN(-ENOMEM);
816
817         lut = lod2lu_dev(lod)->ld_site->ls_tgt;
818         tdtd->tdtd_dt = &lod->lod_dt_dev;
819         rc = distribute_txn_init(env, lut, tdtd,
820                 lu_site2seq(lod2lu_dev(lod)->ld_site)->ss_node_id);
821
822         if (rc < 0) {
823                 CERROR("%s: cannot init distribute txn: rc = %d\n",
824                        lod2obd(lod)->obd_name, rc);
825                 OBD_FREE_PTR(tdtd);
826                 RETURN(rc);
827         }
828
829         tdtd->tdtd_show_update_logs_retrievers =
830                 lod_show_update_logs_retrievers;
831         tdtd->tdtd_show_retrievers_cbdata = lod;
832
833         lut->lut_tdtd = tdtd;
834
835         RETURN(0);
836 }
837
838 /**
839  * Finish distribute txn
840  *
841  * Release the resource holding by distribute txn, i.e. stop distribute
842  * txn thread.
843  *
844  * \param[in] env       execution environment
845  * \param[in] lod       lod device
846  */
847 static void lod_fini_distribute_txn(const struct lu_env *env,
848                                     struct lod_device *lod)
849 {
850         struct lu_target                  *lut;
851
852         lut = lod2lu_dev(lod)->ld_site->ls_tgt;
853         target_recovery_fini(lut->lut_obd);
854         if (lut->lut_tdtd == NULL)
855                 return;
856
857         distribute_txn_fini(env, lut->lut_tdtd);
858
859         OBD_FREE_PTR(lut->lut_tdtd);
860         lut->lut_tdtd = NULL;
861 }
862
863 /**
864  * Implementation of lu_device_operations::ldo_process_config() for LOD
865  *
866  * The method is called by the configuration subsystem during setup,
867  * cleanup and when the configuration changes. The method processes
868  * few specific commands like adding/removing the targets, changing
869  * the runtime parameters.
870
871  * \param[in] env               LU environment provided by the caller
872  * \param[in] dev               lod device
873  * \param[in] lcfg              configuration command to apply
874  *
875  * \retval 0                    on success
876  * \retval negative             negated errno on error
877  *
878  * The examples are below.
879  *
880  * Add osc config log:
881  * marker  20 (flags=0x01, v2.2.49.56) lustre-OST0001  'add osc'
882  * add_uuid  nid=192.168.122.162@tcp(0x20000c0a87aa2)  0:  1:nidxxx
883  * attach    0:lustre-OST0001-osc-MDT0001  1:osc  2:lustre-MDT0001-mdtlov_UUID
884  * setup     0:lustre-OST0001-osc-MDT0001  1:lustre-OST0001_UUID  2:nid
885  * lov_modify_tgts add 0:lustre-MDT0001-mdtlov  1:lustre-OST0001_UUID  2:1  3:1
886  * marker  20 (flags=0x02, v2.2.49.56) lustre-OST0001  'add osc'
887  *
888  * Add mdc config log:
889  * marker  10 (flags=0x01, v2.2.49.56) lustre-MDT0000  'add osp'
890  * add_uuid  nid=192.168.122.162@tcp(0x20000c0a87aa2)  0:  1:nid
891  * attach 0:lustre-MDT0000-osp-MDT0001  1:osp  2:lustre-MDT0001-mdtlov_UUID
892  * setup     0:lustre-MDT0000-osp-MDT0001  1:lustre-MDT0000_UUID  2:nid
893  * modify_mdc_tgts add 0:lustre-MDT0001  1:lustre-MDT0000_UUID  2:0  3:1
894  * marker  10 (flags=0x02, v2.2.49.56) lustre-MDT0000_UUID  'add osp'
895  */
896 static int lod_process_config(const struct lu_env *env,
897                               struct lu_device *dev,
898                               struct lustre_cfg *lcfg)
899 {
900         struct lod_device *lod = lu2lod_dev(dev);
901         struct lu_device  *next = &lod->lod_child->dd_lu_dev;
902         char              *arg1;
903         int                rc = 0;
904         ENTRY;
905
906         switch(lcfg->lcfg_command) {
907         case LCFG_LOV_DEL_OBD:
908         case LCFG_LOV_ADD_INA:
909         case LCFG_LOV_ADD_OBD:
910         case LCFG_ADD_MDC: {
911                 __u32 index;
912                 __u32 mdt_index;
913                 int gen;
914                 /* lov_modify_tgts add  0:lov_mdsA  1:osp  2:0  3:1
915                  * modify_mdc_tgts add  0:lustre-MDT0001
916                  *                    1:lustre-MDT0001-mdc0002
917                  *                    2:2  3:1*/
918                 arg1 = lustre_cfg_string(lcfg, 1);
919
920                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%d", &index) != 1)
921                         GOTO(out, rc = -EINVAL);
922                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
923                         GOTO(out, rc = -EINVAL);
924
925                 if (lcfg->lcfg_command == LCFG_LOV_ADD_OBD) {
926                         __u32 mdt_index;
927
928                         rc = lodname2mdt_index(lustre_cfg_string(lcfg, 0),
929                                                &mdt_index);
930                         if (rc != 0)
931                                 GOTO(out, rc);
932
933                         rc = lod_add_device(env, lod, arg1, index, gen,
934                                             mdt_index, LUSTRE_OSC_NAME, 1);
935                 } else if (lcfg->lcfg_command == LCFG_ADD_MDC) {
936                         mdt_index = index;
937                         rc = lod_add_device(env, lod, arg1, index, gen,
938                                             mdt_index, LUSTRE_MDC_NAME, 1);
939                 } else if (lcfg->lcfg_command == LCFG_LOV_ADD_INA) {
940                         /*FIXME: Add mdt_index for LCFG_LOV_ADD_INA*/
941                         mdt_index = 0;
942                         rc = lod_add_device(env, lod, arg1, index, gen,
943                                             mdt_index, LUSTRE_OSC_NAME, 0);
944                 } else {
945                         rc = lod_del_device(env, lod,
946                                             &lod->lod_ost_descs,
947                                             arg1, index, gen, true);
948                 }
949
950                 break;
951         }
952
953         case LCFG_PARAM: {
954                 struct obd_device *obd;
955                 char *param;
956
957                 /* Check if it is activate/deactivate mdc
958                  * lustre-MDTXXXX-osp-MDTXXXX.active=1 */
959                 param = lustre_cfg_buf(lcfg, 1);
960                 if (strstr(param, "osp") != NULL &&
961                     strstr(param, ".active=") != NULL) {
962                         struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
963                         struct lod_tgt_desc     *sub_tgt = NULL;
964                         char *ptr;
965                         char *tmp;
966                         int i;
967
968                         ptr = strstr(param, ".");
969                         *ptr = '\0';
970                         obd = class_name2obd(param);
971                         if (obd == NULL) {
972                                 CERROR("%s: can not find %s: rc = %d\n",
973                                        lod2obd(lod)->obd_name, param, -EINVAL);
974                                 *ptr = '.';
975                                 GOTO(out, rc);
976                         }
977
978                         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
979                                 struct lod_tgt_desc *tgt;
980
981                                 tgt = LTD_TGT(ltd, i);
982                                 if (tgt->ltd_tgt->dd_lu_dev.ld_obd == obd) {
983                                         sub_tgt = tgt;
984                                         break;
985                                 }
986                         }
987
988                         if (sub_tgt == NULL) {
989                                 CERROR("%s: can not find %s: rc = %d\n",
990                                        lod2obd(lod)->obd_name, param, -EINVAL);
991                                 *ptr = '.';
992                                 GOTO(out, rc);
993                         }
994
995                         *ptr = '.';
996                         tmp = strstr(param, "=");
997                         tmp++;
998                         if (*tmp == '1') {
999                                 struct llog_ctxt *ctxt;
1000
1001                                 obd = sub_tgt->ltd_tgt->dd_lu_dev.ld_obd;
1002                                 ctxt = llog_get_context(obd,
1003                                                 LLOG_UPDATELOG_ORIG_CTXT);
1004                                 if (ctxt == NULL) {
1005                                         rc = llog_setup(env, obd, &obd->obd_olg,
1006                                                        LLOG_UPDATELOG_ORIG_CTXT,
1007                                                     NULL, &llog_common_cat_ops);
1008                                         if (rc < 0)
1009                                                 GOTO(out, rc);
1010                                 } else {
1011                                         llog_ctxt_put(ctxt);
1012                                 }
1013                                 rc = lod_sub_prep_llog(env, lod,
1014                                                        sub_tgt->ltd_tgt,
1015                                                        sub_tgt->ltd_index);
1016                                 if (rc == 0)
1017                                         sub_tgt->ltd_active = 1;
1018                         } else {
1019                                 lod_sub_fini_llog(env, sub_tgt->ltd_tgt,
1020                                                   NULL);
1021                                 sub_tgt->ltd_active = 0;
1022                         }
1023                         GOTO(out, rc);
1024                 }
1025
1026                 obd = lod2obd(lod);
1027                 rc = class_process_proc_param(PARAM_LOV, obd->obd_vars,
1028                                               lcfg, obd);
1029                 if (rc > 0)
1030                         rc = 0;
1031                 GOTO(out, rc);
1032         }
1033         case LCFG_PRE_CLEANUP: {
1034                 if (lod->lod_md_root != NULL) {
1035                         lu_object_put(env, &lod->lod_md_root->ldo_obj.do_lu);
1036                         lod->lod_md_root = NULL;
1037                 }
1038
1039                 lod_sub_process_config(env, lod, &lod->lod_mdt_descs, lcfg);
1040                 lod_sub_process_config(env, lod, &lod->lod_ost_descs, lcfg);
1041                 next = &lod->lod_child->dd_lu_dev;
1042                 rc = next->ld_ops->ldo_process_config(env, next, lcfg);
1043                 if (rc != 0)
1044                         CDEBUG(D_HA, "%s: can't process %u: %d\n",
1045                                lod2obd(lod)->obd_name, lcfg->lcfg_command, rc);
1046
1047                 lod_sub_stop_recovery_threads(env, lod);
1048                 lod_fini_distribute_txn(env, lod);
1049                 lod_sub_fini_all_llogs(env, lod);
1050                 break;
1051         }
1052         case LCFG_CLEANUP: {
1053                 /*
1054                  * do cleanup on underlying storage only when
1055                  * all OSPs are cleaned up, as they use that OSD as well
1056                  */
1057                 lu_dev_del_linkage(dev->ld_site, dev);
1058                 lod_sub_process_config(env, lod, &lod->lod_mdt_descs, lcfg);
1059                 lod_sub_process_config(env, lod, &lod->lod_ost_descs, lcfg);
1060                 next = &lod->lod_child->dd_lu_dev;
1061                 rc = next->ld_ops->ldo_process_config(env, next, lcfg);
1062                 if (rc)
1063                         CERROR("%s: can't process %u: %d\n",
1064                                lod2obd(lod)->obd_name, lcfg->lcfg_command, rc);
1065
1066                 rc = obd_disconnect(lod->lod_child_exp);
1067                 if (rc)
1068                         CERROR("error in disconnect from storage: %d\n", rc);
1069                 break;
1070         }
1071         default:
1072                CERROR("%s: unknown command %u\n", lod2obd(lod)->obd_name,
1073                       lcfg->lcfg_command);
1074                rc = -EINVAL;
1075                break;
1076         }
1077
1078 out:
1079         RETURN(rc);
1080 }
1081
1082 /**
1083  * Implementation of lu_device_operations::ldo_recovery_complete() for LOD
1084  *
1085  * The method is called once the recovery is complete. This implementation
1086  * distributes the notification to all the known targets.
1087  *
1088  * see include/lu_object.h for the details
1089  */
1090 static int lod_recovery_complete(const struct lu_env *env,
1091                                  struct lu_device *dev)
1092 {
1093         struct lod_device   *lod = lu2lod_dev(dev);
1094         struct lu_device    *next = &lod->lod_child->dd_lu_dev;
1095         unsigned int         i;
1096         int                  rc;
1097         ENTRY;
1098
1099         LASSERT(lod->lod_recovery_completed == 0);
1100         lod->lod_recovery_completed = 1;
1101
1102         rc = next->ld_ops->ldo_recovery_complete(env, next);
1103
1104         lod_getref(&lod->lod_ost_descs);
1105         if (lod->lod_osts_size > 0) {
1106                 cfs_foreach_bit(lod->lod_ost_bitmap, i) {
1107                         struct lod_tgt_desc *tgt;
1108                         tgt = OST_TGT(lod, i);
1109                         LASSERT(tgt && tgt->ltd_tgt);
1110                         next = &tgt->ltd_ost->dd_lu_dev;
1111                         rc = next->ld_ops->ldo_recovery_complete(env, next);
1112                         if (rc)
1113                                 CERROR("%s: can't complete recovery on #%d:"
1114                                         "%d\n", lod2obd(lod)->obd_name, i, rc);
1115                 }
1116         }
1117         lod_putref(lod, &lod->lod_ost_descs);
1118         RETURN(rc);
1119 }
1120
1121 /**
1122  * Init update logs on all sub device
1123  *
1124  * LOD initialize update logs on all of sub devices. Because the initialization
1125  * process might need FLD lookup, see llog_osd_open()->dt_locate()->...->
1126  * lod_object_init(), this API has to be called after LOD is initialized.
1127  * \param[in] env       execution environment
1128  * \param[in] lod       lod device
1129  *
1130  * \retval              0 if update log is initialized successfully.
1131  * \retval              negative errno if initialization fails.
1132  */
1133 static int lod_sub_init_llogs(const struct lu_env *env, struct lod_device *lod)
1134 {
1135         struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
1136         int                     rc;
1137         unsigned int            i;
1138         ENTRY;
1139
1140         /* llog must be setup after LOD is initialized, because llog
1141          * initialization include FLD lookup */
1142         LASSERT(lod->lod_initialized);
1143
1144         /* Init the llog in its own stack */
1145         rc = lod_sub_init_llog(env, lod, lod->lod_child);
1146         if (rc < 0)
1147                 RETURN(rc);
1148
1149         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
1150                 struct lod_tgt_desc     *tgt;
1151
1152                 tgt = LTD_TGT(ltd, i);
1153                 rc = lod_sub_init_llog(env, lod, tgt->ltd_tgt);
1154                 if (rc != 0)
1155                         break;
1156         }
1157
1158         RETURN(rc);
1159 }
1160
1161 /**
1162  * Implementation of lu_device_operations::ldo_prepare() for LOD
1163  *
1164  * see include/lu_object.h for the details.
1165  */
1166 static int lod_prepare(const struct lu_env *env, struct lu_device *pdev,
1167                        struct lu_device *cdev)
1168 {
1169         struct lod_device       *lod = lu2lod_dev(cdev);
1170         struct lu_device        *next = &lod->lod_child->dd_lu_dev;
1171         struct lu_fid           *fid = &lod_env_info(env)->lti_fid;
1172         int                     rc;
1173         struct dt_object        *root;
1174         struct dt_object        *dto;
1175         __u32                   index;
1176         ENTRY;
1177
1178         rc = next->ld_ops->ldo_prepare(env, pdev, next);
1179         if (rc != 0) {
1180                 CERROR("%s: prepare bottom error: rc = %d\n",
1181                        lod2obd(lod)->obd_name, rc);
1182                 RETURN(rc);
1183         }
1184
1185         lod->lod_initialized = 1;
1186
1187         rc = dt_root_get(env, lod->lod_child, fid);
1188         if (rc < 0)
1189                 RETURN(rc);
1190
1191         root = dt_locate(env, lod->lod_child, fid);
1192         if (IS_ERR(root))
1193                 RETURN(PTR_ERR(root));
1194
1195         /* Create update log object */
1196         index = lu_site2seq(lod2lu_dev(lod)->ld_site)->ss_node_id;
1197         lu_update_log_fid(fid, index);
1198
1199         dto = local_file_find_or_create_with_fid(env, lod->lod_child,
1200                                                  fid, root,
1201                                                  lod_update_log_name,
1202                                                  S_IFREG | S_IRUGO | S_IWUSR);
1203         if (IS_ERR(dto))
1204                 GOTO(out_put, rc = PTR_ERR(dto));
1205
1206         lu_object_put(env, &dto->do_lu);
1207
1208         /* Create update log dir */
1209         lu_update_log_dir_fid(fid, index);
1210         dto = local_file_find_or_create_with_fid(env, lod->lod_child,
1211                                                  fid, root,
1212                                                  lod_update_log_dir_name,
1213                                                  S_IFDIR | S_IRUGO | S_IWUSR);
1214         if (IS_ERR(dto))
1215                 GOTO(out_put, rc = PTR_ERR(dto));
1216
1217         lu_object_put(env, &dto->do_lu);
1218
1219         rc = lod_prepare_distribute_txn(env, lod);
1220         if (rc != 0)
1221                 GOTO(out_put, rc);
1222
1223         rc = lod_sub_init_llogs(env, lod);
1224         if (rc != 0)
1225                 GOTO(out_put, rc);
1226
1227 out_put:
1228         lu_object_put(env, &root->do_lu);
1229
1230         RETURN(rc);
1231 }
1232
1233 const struct lu_device_operations lod_lu_ops = {
1234         .ldo_object_alloc       = lod_object_alloc,
1235         .ldo_process_config     = lod_process_config,
1236         .ldo_recovery_complete  = lod_recovery_complete,
1237         .ldo_prepare            = lod_prepare,
1238 };
1239
1240 /**
1241  * Implementation of dt_device_operations::dt_root_get() for LOD
1242  *
1243  * see include/dt_object.h for the details.
1244  */
1245 static int lod_root_get(const struct lu_env *env,
1246                         struct dt_device *dev, struct lu_fid *f)
1247 {
1248         return dt_root_get(env, dt2lod_dev(dev)->lod_child, f);
1249 }
1250
1251 /**
1252  * Implementation of dt_device_operations::dt_statfs() for LOD
1253  *
1254  * see include/dt_object.h for the details.
1255  */
1256 static int lod_statfs(const struct lu_env *env,
1257                       struct dt_device *dev, struct obd_statfs *sfs)
1258 {
1259         return dt_statfs(env, dt2lod_dev(dev)->lod_child, sfs);
1260 }
1261
1262 /**
1263  * Implementation of dt_device_operations::dt_trans_create() for LOD
1264  *
1265  * Creates a transaction using local (to this node) OSD.
1266  *
1267  * see include/dt_object.h for the details.
1268  */
1269 static struct thandle *lod_trans_create(const struct lu_env *env,
1270                                         struct dt_device *dt)
1271 {
1272         struct thandle *th;
1273
1274         th = top_trans_create(env, dt2lod_dev(dt)->lod_child);
1275         if (IS_ERR(th))
1276                 return th;
1277
1278         th->th_dev = dt;
1279
1280         return th;
1281 }
1282
1283 /**
1284  * Implementation of dt_device_operations::dt_trans_start() for LOD
1285  *
1286  * Starts the set of local transactions using the targets involved
1287  * in declare phase. Initial support for the distributed transactions.
1288  *
1289  * see include/dt_object.h for the details.
1290  */
1291 static int lod_trans_start(const struct lu_env *env, struct dt_device *dt,
1292                            struct thandle *th)
1293 {
1294         return top_trans_start(env, dt2lod_dev(dt)->lod_child, th);
1295 }
1296
1297 static int lod_trans_cb_add(struct thandle *th,
1298                             struct dt_txn_commit_cb *dcb)
1299 {
1300         struct top_thandle      *top_th = container_of(th, struct top_thandle,
1301                                                        tt_super);
1302         return dt_trans_cb_add(top_th->tt_master_sub_thandle, dcb);
1303 }
1304
1305 /**
1306  * add noop update to the update records
1307  *
1308  * Add noop updates to the update records, which is only used in
1309  * test right now.
1310  *
1311  * \param[in] env       execution environment
1312  * \param[in] dt        dt device of lod
1313  * \param[in] th        thandle
1314  * \param[in] count     the count of update records to be added.
1315  *
1316  * \retval              0 if adding succeeds.
1317  * \retval              negative errno if adding fails.
1318  */
1319 static int lod_add_noop_records(const struct lu_env *env,
1320                                 struct dt_device *dt, struct thandle *th,
1321                                 int count)
1322 {
1323         struct top_thandle *top_th;
1324         struct lu_fid *fid = &lod_env_info(env)->lti_fid;
1325         int i;
1326         int rc = 0;
1327
1328         top_th = container_of(th, struct top_thandle, tt_super);
1329         if (top_th->tt_multiple_thandle == NULL)
1330                 return 0;
1331
1332         fid_zero(fid);
1333         for (i = 0; i < count; i++) {
1334                 rc = update_record_pack(noop, th, fid);
1335                 if (rc < 0)
1336                         return rc;
1337         }
1338         return rc;
1339 }
1340
1341 /**
1342  * Implementation of dt_device_operations::dt_trans_stop() for LOD
1343  *
1344  * Stops the set of local transactions using the targets involved
1345  * in declare phase. Initial support for the distributed transactions.
1346  *
1347  * see include/dt_object.h for the details.
1348  */
1349 static int lod_trans_stop(const struct lu_env *env, struct dt_device *dt,
1350                           struct thandle *th)
1351 {
1352         if (OBD_FAIL_CHECK(OBD_FAIL_SPLIT_UPDATE_REC)) {
1353                 int rc;
1354
1355                 rc = lod_add_noop_records(env, dt, th, 5000);
1356                 if (rc < 0)
1357                         RETURN(rc);
1358         }
1359         return top_trans_stop(env, dt2lod_dev(dt)->lod_child, th);
1360 }
1361
1362 /**
1363  * Implementation of dt_device_operations::dt_conf_get() for LOD
1364  *
1365  * Currently returns the configuration provided by the local OSD.
1366  *
1367  * see include/dt_object.h for the details.
1368  */
1369 static void lod_conf_get(const struct lu_env *env,
1370                          const struct dt_device *dev,
1371                          struct dt_device_param *param)
1372 {
1373         dt_conf_get(env, dt2lod_dev((struct dt_device *)dev)->lod_child, param);
1374 }
1375
1376 /**
1377  * Implementation of dt_device_operations::dt_sync() for LOD
1378  *
1379  * Syncs all known OST targets. Very very expensive and used
1380  * rarely by LFSCK now. Should not be used in general.
1381  *
1382  * see include/dt_object.h for the details.
1383  */
1384 static int lod_sync(const struct lu_env *env, struct dt_device *dev)
1385 {
1386         struct lod_device   *lod = dt2lod_dev(dev);
1387         struct lod_ost_desc *ost;
1388         unsigned int         i;
1389         int                  rc = 0;
1390         ENTRY;
1391
1392         lod_getref(&lod->lod_ost_descs);
1393         lod_foreach_ost(lod, i) {
1394                 ost = OST_TGT(lod, i);
1395                 LASSERT(ost && ost->ltd_ost);
1396                 rc = dt_sync(env, ost->ltd_ost);
1397                 if (rc) {
1398                         CERROR("%s: can't sync %u: %d\n",
1399                                lod2obd(lod)->obd_name, i, rc);
1400                         break;
1401                 }
1402         }
1403         lod_putref(lod, &lod->lod_ost_descs);
1404         if (rc == 0)
1405                 rc = dt_sync(env, lod->lod_child);
1406
1407         RETURN(rc);
1408 }
1409
1410 /**
1411  * Implementation of dt_device_operations::dt_ro() for LOD
1412  *
1413  * Turns local OSD read-only, used for the testing only.
1414  *
1415  * see include/dt_object.h for the details.
1416  */
1417 static int lod_ro(const struct lu_env *env, struct dt_device *dev)
1418 {
1419         return dt_ro(env, dt2lod_dev(dev)->lod_child);
1420 }
1421
1422 /**
1423  * Implementation of dt_device_operations::dt_commit_async() for LOD
1424  *
1425  * Asks local OSD to commit sooner.
1426  *
1427  * see include/dt_object.h for the details.
1428  */
1429 static int lod_commit_async(const struct lu_env *env, struct dt_device *dev)
1430 {
1431         return dt_commit_async(env, dt2lod_dev(dev)->lod_child);
1432 }
1433
1434 static const struct dt_device_operations lod_dt_ops = {
1435         .dt_root_get         = lod_root_get,
1436         .dt_statfs           = lod_statfs,
1437         .dt_trans_create     = lod_trans_create,
1438         .dt_trans_start      = lod_trans_start,
1439         .dt_trans_stop       = lod_trans_stop,
1440         .dt_conf_get         = lod_conf_get,
1441         .dt_sync             = lod_sync,
1442         .dt_ro               = lod_ro,
1443         .dt_commit_async     = lod_commit_async,
1444         .dt_trans_cb_add     = lod_trans_cb_add,
1445 };
1446
1447 /**
1448  * Connect to a local OSD.
1449  *
1450  * Used to connect to the local OSD at mount. OSD name is taken from the
1451  * configuration command passed. This connection is used to identify LU
1452  * site and pin the OSD from early removal.
1453  *
1454  * \param[in] env               LU environment provided by the caller
1455  * \param[in] lod               lod device
1456  * \param[in] cfg               configuration command to apply
1457  *
1458  * \retval 0                    on success
1459  * \retval negative             negated errno on error
1460  **/
1461 static int lod_connect_to_osd(const struct lu_env *env, struct lod_device *lod,
1462                               struct lustre_cfg *cfg)
1463 {
1464         struct obd_connect_data *data = NULL;
1465         struct obd_device       *obd;
1466         char                    *nextdev = NULL, *p, *s;
1467         int                      rc, len = 0;
1468         ENTRY;
1469
1470         LASSERT(cfg);
1471         LASSERT(lod->lod_child_exp == NULL);
1472
1473         /* compatibility hack: we still use old config logs
1474          * which specify LOV, but we need to learn underlying
1475          * OSD device, which is supposed to be:
1476          *  <fsname>-MDTxxxx-osd
1477          *
1478          * 2.x MGS generates lines like the following:
1479          *   #03 (176)lov_setup 0:lustre-MDT0000-mdtlov  1:(struct lov_desc)
1480          * 1.8 MGS generates lines like the following:
1481          *   #03 (168)lov_setup 0:lustre-mdtlov  1:(struct lov_desc)
1482          *
1483          * we use "-MDT" to differentiate 2.x from 1.8 */
1484
1485         if ((p = lustre_cfg_string(cfg, 0)) && strstr(p, "-mdtlov")) {
1486                 len = strlen(p) + 6;
1487                 OBD_ALLOC(nextdev, len);
1488                 if (nextdev == NULL)
1489                         GOTO(out, rc = -ENOMEM);
1490
1491                 strcpy(nextdev, p);
1492                 s = strstr(nextdev, "-mdtlov");
1493                 if (unlikely(s == NULL)) {
1494                         CERROR("unable to parse device name %s\n",
1495                                lustre_cfg_string(cfg, 0));
1496                         GOTO(out, rc = -EINVAL);
1497                 }
1498
1499                 if (strstr(nextdev, "-MDT")) {
1500                         /* 2.x config */
1501                         strcpy(s, "-osd");
1502                 } else {
1503                         /* 1.8 config */
1504                         strcpy(s, "-MDT0000-osd");
1505                 }
1506         } else {
1507                 CERROR("unable to parse device name %s\n",
1508                        lustre_cfg_string(cfg, 0));
1509                 GOTO(out, rc = -EINVAL);
1510         }
1511
1512         OBD_ALLOC_PTR(data);
1513         if (data == NULL)
1514                 GOTO(out, rc = -ENOMEM);
1515
1516         obd = class_name2obd(nextdev);
1517         if (obd == NULL) {
1518                 CERROR("can not locate next device: %s\n", nextdev);
1519                 GOTO(out, rc = -ENOTCONN);
1520         }
1521
1522         data->ocd_connect_flags = OBD_CONNECT_VERSION;
1523         data->ocd_version = LUSTRE_VERSION_CODE;
1524
1525         rc = obd_connect(env, &lod->lod_child_exp, obd, &obd->obd_uuid,
1526                          data, NULL);
1527         if (rc) {
1528                 CERROR("cannot connect to next dev %s (%d)\n", nextdev, rc);
1529                 GOTO(out, rc);
1530         }
1531
1532         lod->lod_dt_dev.dd_lu_dev.ld_site =
1533                 lod->lod_child_exp->exp_obd->obd_lu_dev->ld_site;
1534         LASSERT(lod->lod_dt_dev.dd_lu_dev.ld_site);
1535         lod->lod_child = lu2dt_dev(lod->lod_child_exp->exp_obd->obd_lu_dev);
1536
1537 out:
1538         if (data)
1539                 OBD_FREE_PTR(data);
1540         if (nextdev)
1541                 OBD_FREE(nextdev, len);
1542         RETURN(rc);
1543 }
1544
1545 /**
1546  * Allocate and initialize target table.
1547  *
1548  * A helper function to initialize the target table and allocate
1549  * a bitmap of the available targets.
1550  *
1551  * \param[in] ltd               target's table to initialize
1552  *
1553  * \retval 0                    on success
1554  * \retval negative             negated errno on error
1555  **/
1556 static int lod_tgt_desc_init(struct lod_tgt_descs *ltd)
1557 {
1558         mutex_init(&ltd->ltd_mutex);
1559         init_rwsem(&ltd->ltd_rw_sem);
1560
1561         /* the OST array and bitmap are allocated/grown dynamically as OSTs are
1562          * added to the LOD, see lod_add_device() */
1563         ltd->ltd_tgt_bitmap = CFS_ALLOCATE_BITMAP(32);
1564         if (ltd->ltd_tgt_bitmap == NULL)
1565                 RETURN(-ENOMEM);
1566
1567         ltd->ltd_tgts_size  = 32;
1568         ltd->ltd_tgtnr      = 0;
1569
1570         ltd->ltd_death_row = 0;
1571         ltd->ltd_refcount  = 0;
1572         return 0;
1573 }
1574
1575 /**
1576  * Initialize LOD device at setup.
1577  *
1578  * Initializes the given LOD device using the original configuration command.
1579  * The function initiates a connection to the local OSD and initializes few
1580  * internal structures like pools, target tables, etc.
1581  *
1582  * \param[in] env               LU environment provided by the caller
1583  * \param[in] lod               lod device
1584  * \param[in] ldt               not used
1585  * \param[in] cfg               configuration command
1586  *
1587  * \retval 0                    on success
1588  * \retval negative             negated errno on error
1589  **/
1590 static int lod_init0(const struct lu_env *env, struct lod_device *lod,
1591                      struct lu_device_type *ldt, struct lustre_cfg *cfg)
1592 {
1593         struct dt_device_param ddp;
1594         struct obd_device     *obd;
1595         int                    rc;
1596         ENTRY;
1597
1598         obd = class_name2obd(lustre_cfg_string(cfg, 0));
1599         if (obd == NULL) {
1600                 CERROR("Cannot find obd with name %s\n",
1601                        lustre_cfg_string(cfg, 0));
1602                 RETURN(-ENODEV);
1603         }
1604
1605         obd->obd_lu_dev = &lod->lod_dt_dev.dd_lu_dev;
1606         lod->lod_dt_dev.dd_lu_dev.ld_obd = obd;
1607         lod->lod_dt_dev.dd_lu_dev.ld_ops = &lod_lu_ops;
1608         lod->lod_dt_dev.dd_ops = &lod_dt_ops;
1609
1610         rc = lod_connect_to_osd(env, lod, cfg);
1611         if (rc)
1612                 RETURN(rc);
1613
1614         dt_conf_get(env, &lod->lod_dt_dev, &ddp);
1615         lod->lod_osd_max_easize = ddp.ddp_max_ea_size;
1616
1617         /* setup obd to be used with old lov code */
1618         rc = lod_pools_init(lod, cfg);
1619         if (rc)
1620                 GOTO(out_disconnect, rc);
1621
1622         rc = lod_procfs_init(lod);
1623         if (rc)
1624                 GOTO(out_pools, rc);
1625
1626         spin_lock_init(&lod->lod_lock);
1627         spin_lock_init(&lod->lod_connects_lock);
1628         lod_tgt_desc_init(&lod->lod_mdt_descs);
1629         lod_tgt_desc_init(&lod->lod_ost_descs);
1630
1631         RETURN(0);
1632
1633 out_pools:
1634         lod_pools_fini(lod);
1635 out_disconnect:
1636         obd_disconnect(lod->lod_child_exp);
1637         RETURN(rc);
1638 }
1639
1640 /**
1641  * Implementation of lu_device_type_operations::ldto_device_free() for LOD
1642  *
1643  * Releases the memory allocated for LOD device.
1644  *
1645  * see include/lu_object.h for the details.
1646  */
1647 static struct lu_device *lod_device_free(const struct lu_env *env,
1648                                          struct lu_device *lu)
1649 {
1650         struct lod_device *lod = lu2lod_dev(lu);
1651         struct lu_device  *next = &lod->lod_child->dd_lu_dev;
1652         ENTRY;
1653
1654         LASSERTF(atomic_read(&lu->ld_ref) == 0, "lu is %p\n", lu);
1655         dt_device_fini(&lod->lod_dt_dev);
1656         OBD_FREE_PTR(lod);
1657         RETURN(next);
1658 }
1659
1660 /**
1661  * Implementation of lu_device_type_operations::ldto_device_alloc() for LOD
1662  *
1663  * Allocates LOD device and calls the helpers to initialize it.
1664  *
1665  * see include/lu_object.h for the details.
1666  */
1667 static struct lu_device *lod_device_alloc(const struct lu_env *env,
1668                                           struct lu_device_type *type,
1669                                           struct lustre_cfg *lcfg)
1670 {
1671         struct lod_device *lod;
1672         struct lu_device  *lu_dev;
1673
1674         OBD_ALLOC_PTR(lod);
1675         if (lod == NULL) {
1676                 lu_dev = ERR_PTR(-ENOMEM);
1677         } else {
1678                 int rc;
1679
1680                 lu_dev = lod2lu_dev(lod);
1681                 dt_device_init(&lod->lod_dt_dev, type);
1682                 rc = lod_init0(env, lod, type, lcfg);
1683                 if (rc != 0) {
1684                         lod_device_free(env, lu_dev);
1685                         lu_dev = ERR_PTR(rc);
1686                 }
1687         }
1688
1689         return lu_dev;
1690 }
1691
1692 /**
1693  * Implementation of lu_device_type_operations::ldto_device_fini() for LOD
1694  *
1695  * Releases the internal resources used by LOD device.
1696  *
1697  * see include/lu_object.h for the details.
1698  */
1699 static struct lu_device *lod_device_fini(const struct lu_env *env,
1700                                          struct lu_device *d)
1701 {
1702         struct lod_device *lod = lu2lod_dev(d);
1703         int                rc;
1704         ENTRY;
1705
1706         lod_pools_fini(lod);
1707
1708         lod_procfs_fini(lod);
1709
1710         rc = lod_fini_tgt(env, lod, &lod->lod_ost_descs, true);
1711         if (rc)
1712                 CERROR("%s:can not fini ost descs %d\n",
1713                         lod2obd(lod)->obd_name, rc);
1714
1715         rc = lod_fini_tgt(env, lod, &lod->lod_mdt_descs, false);
1716         if (rc)
1717                 CERROR("%s:can not fini mdt descs %d\n",
1718                         lod2obd(lod)->obd_name, rc);
1719
1720         RETURN(NULL);
1721 }
1722
1723 /**
1724  * Implementation of obd_ops::o_connect() for LOD
1725  *
1726  * Used to track all the users of this specific LOD device,
1727  * so the device stays up until the last user disconnected.
1728  *
1729  * \param[in] env               LU environment provided by the caller
1730  * \param[out] exp              export the caller will be using to access LOD
1731  * \param[in] obd               OBD device representing LOD device
1732  * \param[in] cluuid            unique identifier of the caller
1733  * \param[in] data              not used
1734  * \param[in] localdata         not used
1735  *
1736  * \retval 0                    on success
1737  * \retval negative             negated errno on error
1738  **/
1739 static int lod_obd_connect(const struct lu_env *env, struct obd_export **exp,
1740                            struct obd_device *obd, struct obd_uuid *cluuid,
1741                            struct obd_connect_data *data, void *localdata)
1742 {
1743         struct lod_device    *lod = lu2lod_dev(obd->obd_lu_dev);
1744         struct lustre_handle  conn;
1745         int                   rc;
1746         ENTRY;
1747
1748         CDEBUG(D_CONFIG, "connect #%d\n", lod->lod_connects);
1749
1750         rc = class_connect(&conn, obd, cluuid);
1751         if (rc)
1752                 RETURN(rc);
1753
1754         *exp = class_conn2export(&conn);
1755
1756         spin_lock(&lod->lod_connects_lock);
1757         lod->lod_connects++;
1758         /* at the moment we expect the only user */
1759         LASSERT(lod->lod_connects == 1);
1760         spin_unlock(&lod->lod_connects_lock);
1761
1762         RETURN(0);
1763 }
1764
1765 /**
1766  *
1767  * Implementation of obd_ops::o_disconnect() for LOD
1768  *
1769  * When the caller doesn't need to use this LOD instance, it calls
1770  * obd_disconnect() and LOD releases corresponding export/reference count.
1771  * Once all the users gone, LOD device is released.
1772  *
1773  * \param[in] exp               export provided to the caller in obd_connect()
1774  *
1775  * \retval 0                    on success
1776  * \retval negative             negated errno on error
1777  **/
1778 static int lod_obd_disconnect(struct obd_export *exp)
1779 {
1780         struct obd_device *obd = exp->exp_obd;
1781         struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
1782         int                rc, release = 0;
1783         ENTRY;
1784
1785         /* Only disconnect the underlying layers on the final disconnect. */
1786         spin_lock(&lod->lod_connects_lock);
1787         lod->lod_connects--;
1788         if (lod->lod_connects != 0) {
1789                 /* why should there be more than 1 connect? */
1790                 spin_unlock(&lod->lod_connects_lock);
1791                 CERROR("%s: disconnect #%d\n", exp->exp_obd->obd_name,
1792                        lod->lod_connects);
1793                 goto out;
1794         }
1795         spin_unlock(&lod->lod_connects_lock);
1796
1797         /* the last user of lod has gone, let's release the device */
1798         release = 1;
1799
1800 out:
1801         rc = class_disconnect(exp); /* bz 9811 */
1802
1803         if (rc == 0 && release)
1804                 class_manual_cleanup(obd);
1805         RETURN(rc);
1806 }
1807
1808 LU_KEY_INIT(lod, struct lod_thread_info);
1809
1810 static void lod_key_fini(const struct lu_context *ctx,
1811                 struct lu_context_key *key, void *data)
1812 {
1813         struct lod_thread_info *info = data;
1814         /* allocated in lod_get_lov_ea
1815          * XXX: this is overload, a tread may have such store but used only
1816          * once. Probably better would be pool of such stores per LOD.
1817          */
1818         if (info->lti_ea_store) {
1819                 OBD_FREE_LARGE(info->lti_ea_store, info->lti_ea_store_size);
1820                 info->lti_ea_store = NULL;
1821                 info->lti_ea_store_size = 0;
1822         }
1823         lu_buf_free(&info->lti_linkea_buf);
1824         OBD_FREE_PTR(info);
1825 }
1826
1827 /* context key: lod_thread_key */
1828 LU_CONTEXT_KEY_DEFINE(lod, LCT_MD_THREAD);
1829
1830 LU_TYPE_INIT_FINI(lod, &lod_thread_key);
1831
1832 static struct lu_device_type_operations lod_device_type_ops = {
1833         .ldto_init           = lod_type_init,
1834         .ldto_fini           = lod_type_fini,
1835
1836         .ldto_start          = lod_type_start,
1837         .ldto_stop           = lod_type_stop,
1838
1839         .ldto_device_alloc   = lod_device_alloc,
1840         .ldto_device_free    = lod_device_free,
1841
1842         .ldto_device_fini    = lod_device_fini
1843 };
1844
1845 static struct lu_device_type lod_device_type = {
1846         .ldt_tags     = LU_DEVICE_DT,
1847         .ldt_name     = LUSTRE_LOD_NAME,
1848         .ldt_ops      = &lod_device_type_ops,
1849         .ldt_ctx_tags = LCT_MD_THREAD,
1850 };
1851
1852 /**
1853  * Implementation of obd_ops::o_get_info() for LOD
1854  *
1855  * Currently, there is only one supported key: KEY_OSP_CONNECTED , to provide
1856  * the caller binary status whether LOD has seen connection to any OST target.
1857  * It will also check if the MDT update log context being initialized (if
1858  * needed).
1859  *
1860  * \param[in] env               LU environment provided by the caller
1861  * \param[in] exp               export of the caller
1862  * \param[in] keylen            len of the key
1863  * \param[in] key               the key
1864  * \param[in] vallen            not used
1865  * \param[in] val               not used
1866  *
1867  * \retval                      0 if a connection was seen
1868  * \retval                      -EAGAIN if LOD isn't running yet or no
1869  *                              connection has been seen yet
1870  * \retval                      -EINVAL if not supported key is requested
1871  **/
1872 static int lod_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1873                             __u32 keylen, void *key, __u32 *vallen, void *val)
1874 {
1875         int rc = -EINVAL;
1876
1877         if (KEY_IS(KEY_OSP_CONNECTED)) {
1878                 struct obd_device       *obd = exp->exp_obd;
1879                 struct lod_device       *d;
1880                 struct lod_tgt_desc     *tgt;
1881                 unsigned int            i;
1882                 int                     rc = 1;
1883
1884                 if (!obd->obd_set_up || obd->obd_stopping)
1885                         RETURN(-EAGAIN);
1886
1887                 d = lu2lod_dev(obd->obd_lu_dev);
1888                 lod_getref(&d->lod_ost_descs);
1889                 lod_foreach_ost(d, i) {
1890                         tgt = OST_TGT(d, i);
1891                         LASSERT(tgt && tgt->ltd_tgt);
1892                         rc = obd_get_info(env, tgt->ltd_exp, keylen, key,
1893                                           vallen, val);
1894                         /* one healthy device is enough */
1895                         if (rc == 0)
1896                                 break;
1897                 }
1898                 lod_putref(d, &d->lod_ost_descs);
1899
1900                 lod_getref(&d->lod_mdt_descs);
1901                 lod_foreach_mdt(d, i) {
1902                         struct llog_ctxt *ctxt;
1903
1904                         tgt = MDT_TGT(d, i);
1905                         LASSERT(tgt != NULL);
1906                         LASSERT(tgt->ltd_tgt != NULL);
1907                         if (!tgt->ltd_active)
1908                                 continue;
1909
1910                         ctxt = llog_get_context(tgt->ltd_tgt->dd_lu_dev.ld_obd,
1911                                                 LLOG_UPDATELOG_ORIG_CTXT);
1912                         if (ctxt == NULL) {
1913                                 CDEBUG(D_INFO, "%s: %s is not ready.\n",
1914                                        obd->obd_name,
1915                                       tgt->ltd_tgt->dd_lu_dev.ld_obd->obd_name);
1916                                 rc = -EAGAIN;
1917                                 break;
1918                         }
1919                         if (ctxt->loc_handle == NULL) {
1920                                 CDEBUG(D_INFO, "%s: %s is not ready.\n",
1921                                        obd->obd_name,
1922                                       tgt->ltd_tgt->dd_lu_dev.ld_obd->obd_name);
1923                                 rc = -EAGAIN;
1924                                 llog_ctxt_put(ctxt);
1925                                 break;
1926                         }
1927                         llog_ctxt_put(ctxt);
1928                 }
1929                 lod_putref(d, &d->lod_mdt_descs);
1930
1931                 RETURN(rc);
1932         }
1933
1934         RETURN(rc);
1935 }
1936
1937 static int lod_obd_set_info_async(const struct lu_env *env,
1938                                   struct obd_export *exp,
1939                                   __u32 keylen, void *key,
1940                                   __u32 vallen, void *val,
1941                                   struct ptlrpc_request_set *set)
1942 {
1943         struct obd_device *obd = class_exp2obd(exp);
1944         struct lod_device *d;
1945         struct lod_tgt_desc *tgt;
1946         int no_set = 0;
1947         int i, rc = 0, rc2;
1948         ENTRY;
1949
1950         if (set == NULL) {
1951                 no_set = 1;
1952                 set = ptlrpc_prep_set();
1953                 if (!set)
1954                         RETURN(-ENOMEM);
1955         }
1956
1957         d = lu2lod_dev(obd->obd_lu_dev);
1958         lod_getref(&d->lod_ost_descs);
1959         lod_foreach_ost(d, i) {
1960                 tgt = OST_TGT(d, i);
1961                 LASSERT(tgt && tgt->ltd_tgt);
1962                 if (!tgt->ltd_active)
1963                         continue;
1964
1965                 rc2 = obd_set_info_async(env, tgt->ltd_exp, keylen, key,
1966                                          vallen, val, set);
1967                 if (rc2 != 0 && rc == 0)
1968                         rc = rc2;
1969         }
1970         lod_putref(d, &d->lod_ost_descs);
1971
1972         lod_getref(&d->lod_mdt_descs);
1973         lod_foreach_mdt(d, i) {
1974                 tgt = MDT_TGT(d, i);
1975                 LASSERT(tgt && tgt->ltd_tgt);
1976                 if (!tgt->ltd_active)
1977                         continue;
1978                 rc2 = obd_set_info_async(env, tgt->ltd_exp, keylen, key,
1979                                          vallen, val, set);
1980                 if (rc2 != 0 && rc == 0)
1981                         rc = rc2;
1982         }
1983         lod_putref(d, &d->lod_mdt_descs);
1984
1985
1986         if (no_set) {
1987                 rc2 = ptlrpc_set_wait(set);
1988                 if (rc2 == 0 && rc == 0)
1989                         rc = rc2;
1990                 ptlrpc_set_destroy(set);
1991         }
1992         RETURN(rc);
1993 }
1994
1995 static struct obd_ops lod_obd_device_ops = {
1996         .o_owner        = THIS_MODULE,
1997         .o_connect      = lod_obd_connect,
1998         .o_disconnect   = lod_obd_disconnect,
1999         .o_get_info     = lod_obd_get_info,
2000         .o_set_info_async = lod_obd_set_info_async,
2001         .o_pool_new     = lod_pool_new,
2002         .o_pool_rem     = lod_pool_remove,
2003         .o_pool_add     = lod_pool_add,
2004         .o_pool_del     = lod_pool_del,
2005 };
2006
2007 static int __init lod_init(void)
2008 {
2009         struct obd_type *type;
2010         int rc;
2011
2012         rc = lu_kmem_init(lod_caches);
2013         if (rc)
2014                 return rc;
2015
2016         rc = class_register_type(&lod_obd_device_ops, NULL, true, NULL,
2017                                  LUSTRE_LOD_NAME, &lod_device_type);
2018         if (rc) {
2019                 lu_kmem_fini(lod_caches);
2020                 return rc;
2021         }
2022
2023         /* create "lov" entry in procfs for compatibility purposes */
2024         type = class_search_type(LUSTRE_LOV_NAME);
2025         if (type != NULL && type->typ_procroot != NULL)
2026                 return rc;
2027
2028         type = class_search_type(LUSTRE_LOD_NAME);
2029         type->typ_procsym = lprocfs_register("lov", proc_lustre_root,
2030                                              NULL, NULL);
2031         if (IS_ERR(type->typ_procsym)) {
2032                 CERROR("lod: can't create compat entry \"lov\": %d\n",
2033                        (int)PTR_ERR(type->typ_procsym));
2034                 type->typ_procsym = NULL;
2035         }
2036         return rc;
2037 }
2038
2039 static void __exit lod_exit(void)
2040 {
2041         class_unregister_type(LUSTRE_LOD_NAME);
2042         lu_kmem_fini(lod_caches);
2043 }
2044
2045 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2046 MODULE_DESCRIPTION("Lustre Logical Object Device ("LUSTRE_LOD_NAME")");
2047 MODULE_VERSION(LUSTRE_VERSION_STRING);
2048 MODULE_LICENSE("GPL");
2049
2050 module_init(lod_init);
2051 module_exit(lod_exit);