Whamcloud - gitweb
f640e0e03fa72bf75e3bdf473a2086d91b4f35c5
[fs/lustre-release.git] / lustre / mgc / mgc_request.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.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 2017, 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/mgc/mgc_request.c
33  *
34  * Author: Nathan Rutman <nathan@clusterfs.com>
35  */
36
37 #define DEBUG_SUBSYSTEM S_MGC
38 #define D_MGC D_CONFIG /*|D_WARNING*/
39
40 #include <linux/module.h>
41 #include <linux/kthread.h>
42
43 #include <dt_object.h>
44 #include <lprocfs_status.h>
45 #include <lustre_dlm.h>
46 #include <lustre_disk.h>
47 #include <lustre_log.h>
48 #include <lustre_nodemap.h>
49 #include <lustre_swab.h>
50 #include <obd_class.h>
51 #include <lustre_barrier.h>
52
53 #include "mgc_internal.h"
54
55 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
56                           int type)
57 {
58         __u64 resname = 0;
59
60         if (len > sizeof(resname)) {
61                 CERROR("name too long: %s\n", name);
62                 return -EINVAL;
63         }
64         if (len <= 0) {
65                 CERROR("missing name: %s\n", name);
66                 return -EINVAL;
67         }
68         memcpy(&resname, name, len);
69
70         /* Always use the same endianness for the resid */
71         memset(res_id, 0, sizeof(*res_id));
72         res_id->name[0] = cpu_to_le64(resname);
73         /* XXX: unfortunately, sptlprc and config llog share one lock */
74         switch(type) {
75         case CONFIG_T_CONFIG:
76         case CONFIG_T_SPTLRPC:
77                 resname = 0;
78                 break;
79         case CONFIG_T_RECOVER:
80         case CONFIG_T_PARAMS:
81         case CONFIG_T_NODEMAP:
82         case CONFIG_T_BARRIER:
83                 resname = type;
84                 break;
85         default:
86                 LBUG();
87         }
88         res_id->name[1] = cpu_to_le64(resname);
89         CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
90                res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
91         return 0;
92 }
93
94 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
95 {
96         /* fsname is at most 8 chars long, maybe contain "-".
97          * e.g. "lustre", "SUN-000" */
98         return mgc_name2resid(fsname, strlen(fsname), res_id, type);
99 }
100 EXPORT_SYMBOL(mgc_fsname2resid);
101
102 int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
103 {
104         char *name_end;
105         int len;
106
107         /* logname consists of "fsname-nodetype".
108          * e.g. "lustre-MDT0001", "SUN-000-client"
109          * there is an exception: llog "params" */
110         name_end = strrchr(logname, '-');
111         if (!name_end)
112                 len = strlen(logname);
113         else
114                 len = name_end - logname;
115         return mgc_name2resid(logname, len, res_id, type);
116 }
117 EXPORT_SYMBOL(mgc_logname2resid);
118
119 /********************** config llog list **********************/
120 static struct list_head config_llog_list = LIST_HEAD_INIT(config_llog_list);
121 static DEFINE_SPINLOCK(config_list_lock);       /* protects config_llog_list */
122
123 /* Take a reference to a config log */
124 static int config_log_get(struct config_llog_data *cld)
125 {
126         ENTRY;
127         atomic_inc(&cld->cld_refcount);
128         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
129                 atomic_read(&cld->cld_refcount));
130         RETURN(0);
131 }
132
133 /* Drop a reference to a config log.  When no longer referenced,
134    we can free the config log data */
135 static void config_log_put(struct config_llog_data *cld)
136 {
137         ENTRY;
138
139         if (unlikely(!cld))
140                 RETURN_EXIT;
141
142         CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
143                 atomic_read(&cld->cld_refcount));
144         LASSERT(atomic_read(&cld->cld_refcount) > 0);
145
146         /* spinlock to make sure no item with 0 refcount in the list */
147         if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
148                 list_del(&cld->cld_list_chain);
149                 spin_unlock(&config_list_lock);
150
151                 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
152
153                 config_log_put(cld->cld_barrier);
154                 config_log_put(cld->cld_recover);
155                 config_log_put(cld->cld_params);
156                 config_log_put(cld->cld_nodemap);
157                 config_log_put(cld->cld_sptlrpc);
158                 if (cld_is_sptlrpc(cld))
159                         sptlrpc_conf_log_stop(cld->cld_logname);
160
161                 class_export_put(cld->cld_mgcexp);
162                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
163         }
164
165         EXIT;
166 }
167
168 /* Find a config log by name */
169 static
170 struct config_llog_data *config_log_find(char *logname,
171                                          struct config_llog_instance *cfg)
172 {
173         struct config_llog_data *cld;
174         struct config_llog_data *found = NULL;
175         void *                   instance;
176         ENTRY;
177
178         LASSERT(logname != NULL);
179
180         instance = cfg ? cfg->cfg_instance : NULL;
181         spin_lock(&config_list_lock);
182         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
183                 /* check if instance equals */
184                 if (instance != cld->cld_cfg.cfg_instance)
185                         continue;
186
187                 /* instance may be NULL, should check name */
188                 if (strcmp(logname, cld->cld_logname) == 0) {
189                         found = cld;
190                         config_log_get(found);
191                         break;
192                 }
193         }
194         spin_unlock(&config_list_lock);
195         RETURN(found);
196 }
197
198 static
199 struct config_llog_data *do_config_log_add(struct obd_device *obd,
200                                            char *logname,
201                                            int type,
202                                            struct config_llog_instance *cfg,
203                                            struct super_block *sb)
204 {
205         struct config_llog_data *cld;
206         int rc;
207
208         ENTRY;
209
210         CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
211                cfg ? cfg->cfg_instance : NULL);
212
213         OBD_ALLOC(cld, sizeof(*cld) + strlen(logname) + 1);
214         if (!cld)
215                 RETURN(ERR_PTR(-ENOMEM));
216
217         rc = mgc_logname2resid(logname, &cld->cld_resid, type);
218         if (rc) {
219                 OBD_FREE(cld, sizeof(*cld) + strlen(cld->cld_logname) + 1);
220                 RETURN(ERR_PTR(rc));
221         }
222
223         strcpy(cld->cld_logname, logname);
224         if (cfg)
225                 cld->cld_cfg = *cfg;
226         else
227                 cld->cld_cfg.cfg_callback = class_config_llog_handler;
228         mutex_init(&cld->cld_lock);
229         cld->cld_cfg.cfg_last_idx = 0;
230         cld->cld_cfg.cfg_flags = 0;
231         cld->cld_cfg.cfg_sb = sb;
232         cld->cld_type = type;
233         atomic_set(&cld->cld_refcount, 1);
234
235         /* Keep the mgc around until we are done */
236         cld->cld_mgcexp = class_export_get(obd->obd_self_export);
237
238         if (cld_is_sptlrpc(cld)) {
239                 sptlrpc_conf_log_start(logname);
240                 cld->cld_cfg.cfg_obdname = obd->obd_name;
241         }
242
243         spin_lock(&config_list_lock);
244         list_add(&cld->cld_list_chain, &config_llog_list);
245         spin_unlock(&config_list_lock);
246
247         if (cld_is_sptlrpc(cld) || cld_is_nodemap(cld) || cld_is_barrier(cld)) {
248                 rc = mgc_process_log(obd, cld);
249                 if (rc && rc != -ENOENT)
250                         CERROR("%s: failed processing log, type %d: rc = %d\n",
251                                obd->obd_name, type, rc);
252         }
253
254         RETURN(cld);
255 }
256
257 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
258         char *fsname,
259         struct config_llog_instance *cfg,
260         struct super_block *sb)
261 {
262         struct config_llog_instance lcfg = *cfg;
263         struct lustre_sb_info *lsi = s2lsi(sb);
264         struct config_llog_data *cld;
265         char logname[32];
266
267         if (IS_OST(lsi))
268                 return NULL;
269
270         /* for osp-on-ost, see lustre_start_osp() */
271         if (IS_MDT(lsi) && lcfg.cfg_instance)
272                 return NULL;
273
274         /* we have to use different llog for clients and mdts for cmd
275          * where only clients are notified if one of cmd server restarts */
276         LASSERT(strlen(fsname) < sizeof(logname) / 2);
277         strcpy(logname, fsname);
278         if (IS_SERVER(lsi)) { /* mdt */
279                 LASSERT(lcfg.cfg_instance == NULL);
280                 lcfg.cfg_instance = sb;
281                 strcat(logname, "-mdtir");
282         } else {
283                 LASSERT(lcfg.cfg_instance != NULL);
284                 strcat(logname, "-cliir");
285         }
286
287         cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
288         return cld;
289 }
290
291 static struct config_llog_data *config_log_find_or_add(struct obd_device *obd,
292                                 char *logname, struct super_block *sb, int type,
293                                 struct config_llog_instance *cfg)
294 {
295         struct config_llog_instance     lcfg = *cfg;
296         struct config_llog_data         *cld;
297
298         lcfg.cfg_instance = sb != NULL ? (void *)sb : (void *)obd;
299
300         if (type == CONFIG_T_SPTLRPC)
301                 lcfg.cfg_instance = NULL;
302
303         cld = config_log_find(logname, &lcfg);
304         if (unlikely(cld != NULL))
305                 return cld;
306
307         return do_config_log_add(obd, logname, type, &lcfg, sb);
308 }
309
310 /** Add this log to the list of active logs watched by an MGC.
311  * Active means we're watching for updates.
312  * We have one active log per "mount" - client instance or servername.
313  * Each instance may be at a different point in the log.
314  */
315 static struct config_llog_data *
316 config_log_add(struct obd_device *obd, char *logname,
317                struct config_llog_instance *cfg, struct super_block *sb)
318 {
319         struct lustre_sb_info *lsi = s2lsi(sb);
320         struct config_llog_data *cld = NULL;
321         struct config_llog_data *sptlrpc_cld = NULL;
322         struct config_llog_data *params_cld = NULL;
323         struct config_llog_data *nodemap_cld = NULL;
324         struct config_llog_data *barrier_cld = NULL;
325         char seclogname[32];
326         char *ptr;
327         int rc;
328         bool locked = false;
329         ENTRY;
330
331         CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
332
333         /*
334          * for each regular log, the depended sptlrpc log name is
335          * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
336          */
337         ptr = strrchr(logname, '-');
338         if (ptr == NULL || ptr - logname > 8) {
339                 CERROR("logname %s is too long\n", logname);
340                 RETURN(ERR_PTR(-EINVAL));
341         }
342
343         memcpy(seclogname, logname, ptr - logname);
344         strcpy(seclogname + (ptr - logname), "-sptlrpc");
345
346         if (cfg->cfg_sub_clds & CONFIG_SUB_SPTLRPC) {
347                 sptlrpc_cld = config_log_find_or_add(obd, seclogname, NULL,
348                                                      CONFIG_T_SPTLRPC, cfg);
349                 if (IS_ERR(sptlrpc_cld)) {
350                         CERROR("%s: can't create sptlrpc log %s: rc = %ld\n",
351                                obd->obd_name, seclogname, PTR_ERR(sptlrpc_cld));
352                         RETURN(sptlrpc_cld);
353                 }
354         }
355
356         if (!IS_MGS(lsi) && cfg->cfg_sub_clds & CONFIG_SUB_NODEMAP) {
357                 nodemap_cld = config_log_find_or_add(obd, LUSTRE_NODEMAP_NAME,
358                                                      NULL, CONFIG_T_NODEMAP,
359                                                      cfg);
360                 if (IS_ERR(nodemap_cld)) {
361                         rc = PTR_ERR(nodemap_cld);
362                         CERROR("%s: cannot create nodemap log: rc = %d\n",
363                                obd->obd_name, rc);
364                         GOTO(out_sptlrpc, rc);
365                 }
366         }
367
368         if (cfg->cfg_sub_clds & CONFIG_SUB_PARAMS) {
369                 params_cld = config_log_find_or_add(obd, PARAMS_FILENAME, sb,
370                                                     CONFIG_T_PARAMS, cfg);
371                 if (IS_ERR(params_cld)) {
372                         rc = PTR_ERR(params_cld);
373                         CERROR("%s: can't create params log: rc = %d\n",
374                                obd->obd_name, rc);
375                         GOTO(out_nodemap, rc);
376                 }
377         }
378
379         if (IS_MDT(s2lsi(sb)) && cfg->cfg_sub_clds & CONFIG_SUB_BARRIER) {
380                 snprintf(seclogname + (ptr - logname), sizeof(seclogname) - 1,
381                          "-%s", BARRIER_FILENAME);
382                 barrier_cld = config_log_find_or_add(obd, seclogname, sb,
383                                                      CONFIG_T_BARRIER, cfg);
384                 if (IS_ERR(barrier_cld)) {
385                         rc = PTR_ERR(barrier_cld);
386                         CERROR("%s: can't create barrier log: rc = %d\n",
387                                obd->obd_name, rc);
388                         GOTO(out_params, rc);
389                 }
390         }
391
392         cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
393         if (IS_ERR(cld)) {
394                 rc = PTR_ERR(cld);
395                 CERROR("%s: can't create log: rc = %d\n",
396                        obd->obd_name, rc);
397                 GOTO(out_barrier, rc = PTR_ERR(cld));
398         }
399
400         LASSERT(lsi->lsi_lmd);
401         if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR) &&
402             cfg->cfg_sub_clds & CONFIG_SUB_RECOVER) {
403                 struct config_llog_data *recover_cld;
404
405                 ptr = strrchr(seclogname, '-');
406                 if (ptr != NULL) {
407                         *ptr = 0;
408                 } else {
409                         CERROR("%s: sptlrpc log name not correct, %s: "
410                                "rc = %d\n", obd->obd_name, seclogname, -EINVAL);
411                         GOTO(out_cld, rc = -EINVAL);
412                 }
413
414                 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
415                 if (IS_ERR(recover_cld)) {
416                         rc = PTR_ERR(recover_cld);
417                         CERROR("%s: can't create recover log: rc = %d\n",
418                                obd->obd_name, rc);
419                         GOTO(out_cld, rc);
420                 }
421
422                 mutex_lock(&cld->cld_lock);
423                 locked = true;
424                 cld->cld_recover = recover_cld;
425         }
426
427         if (!locked)
428                 mutex_lock(&cld->cld_lock);
429         cld->cld_params = params_cld;
430         cld->cld_barrier = barrier_cld;
431         cld->cld_nodemap = nodemap_cld;
432         cld->cld_sptlrpc = sptlrpc_cld;
433         mutex_unlock(&cld->cld_lock);
434
435         RETURN(cld);
436
437 out_cld:
438         config_log_put(cld);
439 out_barrier:
440         config_log_put(barrier_cld);
441 out_params:
442         config_log_put(params_cld);
443 out_nodemap:
444         config_log_put(nodemap_cld);
445 out_sptlrpc:
446         config_log_put(sptlrpc_cld);
447
448         return ERR_PTR(rc);
449 }
450
451 DEFINE_MUTEX(llog_process_lock);
452
453 static inline void config_mark_cld_stop(struct config_llog_data *cld)
454 {
455         if (cld) {
456                 mutex_lock(&cld->cld_lock);
457                 spin_lock(&config_list_lock);
458                 cld->cld_stopping = 1;
459                 spin_unlock(&config_list_lock);
460                 mutex_unlock(&cld->cld_lock);
461         }
462 }
463
464 /** Stop watching for updates on this log.
465  */
466 static int config_log_end(char *logname, struct config_llog_instance *cfg)
467 {
468         struct config_llog_data *cld;
469         struct config_llog_data *cld_sptlrpc = NULL;
470         struct config_llog_data *cld_params = NULL;
471         struct config_llog_data *cld_recover = NULL;
472         struct config_llog_data *cld_nodemap = NULL;
473         struct config_llog_data *cld_barrier = NULL;
474         int rc = 0;
475
476         ENTRY;
477
478         cld = config_log_find(logname, cfg);
479         if (cld == NULL)
480                 RETURN(-ENOENT);
481
482         mutex_lock(&cld->cld_lock);
483         /*
484          * if cld_stopping is set, it means we didn't start the log thus
485          * not owning the start ref. this can happen after previous umount:
486          * the cld still hanging there waiting for lock cancel, and we
487          * remount again but failed in the middle and call log_end without
488          * calling start_log.
489          */
490         if (unlikely(cld->cld_stopping)) {
491                 mutex_unlock(&cld->cld_lock);
492                 /* drop the ref from the find */
493                 config_log_put(cld);
494                 RETURN(rc);
495         }
496
497         spin_lock(&config_list_lock);
498         cld->cld_stopping = 1;
499         spin_unlock(&config_list_lock);
500
501         cld_recover = cld->cld_recover;
502         cld->cld_recover = NULL;
503         cld_params = cld->cld_params;
504         cld->cld_params = NULL;
505         cld_nodemap = cld->cld_nodemap;
506         cld->cld_nodemap = NULL;
507         cld_barrier = cld->cld_barrier;
508         cld->cld_barrier = NULL;
509         cld_sptlrpc = cld->cld_sptlrpc;
510         cld->cld_sptlrpc = NULL;
511         mutex_unlock(&cld->cld_lock);
512
513         config_mark_cld_stop(cld_recover);
514         config_log_put(cld_recover);
515
516         config_mark_cld_stop(cld_params);
517         config_log_put(cld_params);
518
519         /* don't set cld_stopping on nm lock as other targets may be active */
520         config_log_put(cld_nodemap);
521
522         if (cld_barrier) {
523                 mutex_lock(&cld_barrier->cld_lock);
524                 cld_barrier->cld_stopping = 1;
525                 mutex_unlock(&cld_barrier->cld_lock);
526                 config_log_put(cld_barrier);
527         }
528
529         config_log_put(cld_sptlrpc);
530
531         /* drop the ref from the find */
532         config_log_put(cld);
533         /* drop the start ref */
534         config_log_put(cld);
535
536         CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
537                rc);
538         RETURN(rc);
539 }
540
541 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
542 {
543         struct obd_device       *obd = data;
544         struct obd_import       *imp;
545         struct obd_connect_data *ocd;
546         struct config_llog_data *cld;
547
548         ENTRY;
549         LASSERT(obd);
550         LPROCFS_CLIMP_CHECK(obd);
551         imp = obd->u.cli.cl_import;
552         ocd = &imp->imp_connect_data;
553
554         seq_printf(m, "imperative_recovery: %s\n",
555                    OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
556         seq_printf(m, "client_state:\n");
557
558         spin_lock(&config_list_lock);
559         list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
560                 if (cld->cld_recover == NULL)
561                         continue;
562                 seq_printf(m,  "    - { client: %s, nidtbl_version: %u }\n",
563                            cld->cld_logname,
564                            cld->cld_recover->cld_cfg.cfg_last_idx);
565         }
566         spin_unlock(&config_list_lock);
567
568         LPROCFS_CLIMP_EXIT(obd);
569         RETURN(0);
570 }
571
572 /* reenqueue any lost locks */
573 #define RQ_RUNNING      0x1
574 #define RQ_NOW          0x2
575 #define RQ_LATER        0x4
576 #define RQ_STOP         0x8
577 #define RQ_PRECLEANUP   0x10
578 static int                    rq_state = 0;
579 static wait_queue_head_t      rq_waitq;
580 static DECLARE_COMPLETION(rq_exit);
581 static DECLARE_COMPLETION(rq_start);
582
583 static void do_requeue(struct config_llog_data *cld)
584 {
585         int rc = 0;
586         ENTRY;
587
588         LASSERT(atomic_read(&cld->cld_refcount) > 0);
589
590         /*
591          * Do not run mgc_process_log on a disconnected export or an
592          * export which is being disconnected. Take the client
593          * semaphore to make the check non-racy.
594          */
595         down_read_nested(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem,
596                          OBD_CLI_SEM_MGC);
597         if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
598                 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
599                 rc = mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
600                 if (rc && rc != -ENOENT)
601                         CERROR("failed processing log: %d\n", rc);
602         } else {
603                 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
604                        cld->cld_logname);
605         }
606         up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
607
608         EXIT;
609 }
610
611 /* this timeout represents how many seconds MGC should wait before
612  * requeue config and recover lock to the MGS. We need to randomize this
613  * in order to not flood the MGS.
614  */
615 #define MGC_TIMEOUT_MIN_SECONDS   5
616 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
617
618 static int mgc_requeue_thread(void *data)
619 {
620         int rc = 0;
621         bool first = true;
622         ENTRY;
623
624         CDEBUG(D_MGC, "Starting requeue thread\n");
625
626         /* Keep trying failed locks periodically */
627         spin_lock(&config_list_lock);
628         rq_state |= RQ_RUNNING;
629         while (!(rq_state & RQ_STOP)) {
630                 struct l_wait_info lwi;
631                 struct config_llog_data *cld, *cld_prev;
632                 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
633                 int to;
634
635                 /* Any new or requeued lostlocks will change the state */
636                 rq_state &= ~(RQ_NOW | RQ_LATER);
637                 spin_unlock(&config_list_lock);
638
639                 if (first) {
640                         first = false;
641                         complete(&rq_start);
642                 }
643
644                 /* Always wait a few seconds to allow the server who
645                    caused the lock revocation to finish its setup, plus some
646                    random so everyone doesn't try to reconnect at once. */
647                 to = msecs_to_jiffies(MGC_TIMEOUT_MIN_SECONDS * MSEC_PER_SEC);
648                 /* rand is centi-seconds */
649                 to += msecs_to_jiffies(rand * MSEC_PER_SEC / 100);
650                 lwi = LWI_TIMEOUT(to, NULL, NULL);
651                 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
652                              &lwi);
653
654                 /*
655                  * iterate & processing through the list. for each cld, process
656                  * its depending sptlrpc cld firstly (if any) and then itself.
657                  *
658                  * it's guaranteed any item in the list must have
659                  * reference > 0; and if cld_lostlock is set, at
660                  * least one reference is taken by the previous enqueue.
661                  */
662                 cld_prev = NULL;
663
664                 spin_lock(&config_list_lock);
665                 rq_state &= ~RQ_PRECLEANUP;
666                 list_for_each_entry(cld, &config_llog_list,
667                                     cld_list_chain) {
668                         if (!cld->cld_lostlock || cld->cld_stopping)
669                                 continue;
670
671                         /* hold reference to avoid being freed during
672                          * subsequent processing. */
673                         config_log_get(cld);
674                         cld->cld_lostlock = 0;
675                         spin_unlock(&config_list_lock);
676
677                         config_log_put(cld_prev);
678                         cld_prev = cld;
679
680                         if (likely(!(rq_state & RQ_STOP))) {
681                                 do_requeue(cld);
682                                 spin_lock(&config_list_lock);
683                         } else {
684                                 spin_lock(&config_list_lock);
685                                 break;
686                         }
687                 }
688                 spin_unlock(&config_list_lock);
689                 config_log_put(cld_prev);
690
691                 /* Wait a bit to see if anyone else needs a requeue */
692                 lwi = (struct l_wait_info) { 0 };
693                 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
694                              &lwi);
695                 spin_lock(&config_list_lock);
696         }
697
698         /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
699         rq_state &= ~RQ_RUNNING;
700         spin_unlock(&config_list_lock);
701
702         complete(&rq_exit);
703
704         CDEBUG(D_MGC, "Ending requeue thread\n");
705         RETURN(rc);
706 }
707
708 /* Add a cld to the list to requeue.  Start the requeue thread if needed.
709    We are responsible for dropping the config log reference from here on out. */
710 static void mgc_requeue_add(struct config_llog_data *cld)
711 {
712         bool wakeup = false;
713         ENTRY;
714
715         CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
716                 cld->cld_logname, atomic_read(&cld->cld_refcount),
717                 cld->cld_stopping, rq_state);
718         LASSERT(atomic_read(&cld->cld_refcount) > 0);
719
720         mutex_lock(&cld->cld_lock);
721         spin_lock(&config_list_lock);
722         if (!(rq_state & RQ_STOP) && !cld->cld_stopping && !cld->cld_lostlock) {
723                 cld->cld_lostlock = 1;
724                 rq_state |= RQ_NOW;
725                 wakeup = true;
726         }
727         spin_unlock(&config_list_lock);
728         mutex_unlock(&cld->cld_lock);
729         if (wakeup)
730                 wake_up(&rq_waitq);
731
732         EXIT;
733 }
734
735 /********************** class fns **********************/
736 static int mgc_local_llog_init(const struct lu_env *env,
737                                struct obd_device *obd,
738                                struct obd_device *disk)
739 {
740         struct llog_ctxt        *ctxt;
741         int                      rc;
742
743         ENTRY;
744
745         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_ORIG_CTXT, disk,
746                         &llog_osd_ops);
747         if (rc)
748                 RETURN(rc);
749
750         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
751         LASSERT(ctxt);
752         ctxt->loc_dir = obd->u.cli.cl_mgc_configs_dir;
753         llog_ctxt_put(ctxt);
754
755         RETURN(0);
756 }
757
758 static int mgc_local_llog_fini(const struct lu_env *env,
759                                struct obd_device *obd)
760 {
761         struct llog_ctxt *ctxt;
762
763         ENTRY;
764
765         ctxt = llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT);
766         llog_cleanup(env, ctxt);
767
768         RETURN(0);
769 }
770
771 static int mgc_fs_setup(const struct lu_env *env, struct obd_device *obd,
772                         struct super_block *sb)
773 {
774         struct lustre_sb_info   *lsi = s2lsi(sb);
775         struct client_obd       *cli = &obd->u.cli;
776         struct lu_fid            rfid, fid;
777         struct dt_object        *root, *dto;
778         int                      rc = 0;
779
780         ENTRY;
781
782         LASSERT(lsi);
783         LASSERT(lsi->lsi_dt_dev);
784
785         /* The mgc fs exclusion mutex. Only one fs can be setup at a time. */
786         mutex_lock(&cli->cl_mgc_mutex);
787
788         /* Setup the configs dir */
789         fid.f_seq = FID_SEQ_LOCAL_NAME;
790         fid.f_oid = 1;
791         fid.f_ver = 0;
792         rc = local_oid_storage_init(env, lsi->lsi_dt_dev, &fid,
793                                     &cli->cl_mgc_los);
794         if (rc)
795                 RETURN(rc);
796
797         rc = dt_root_get(env, lsi->lsi_dt_dev, &rfid);
798         if (rc)
799                 GOTO(out_los, rc);
800
801         root = dt_locate_at(env, lsi->lsi_dt_dev, &rfid,
802                             &cli->cl_mgc_los->los_dev->dd_lu_dev, NULL);
803         if (unlikely(IS_ERR(root)))
804                 GOTO(out_los, rc = PTR_ERR(root));
805
806         dto = local_file_find_or_create(env, cli->cl_mgc_los, root,
807                                         MOUNT_CONFIGS_DIR,
808                                         S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO);
809         dt_object_put_nocache(env, root);
810         if (IS_ERR(dto))
811                 GOTO(out_los, rc = PTR_ERR(dto));
812
813         cli->cl_mgc_configs_dir = dto;
814
815         LASSERT(lsi->lsi_osd_exp->exp_obd->obd_lvfs_ctxt.dt);
816         rc = mgc_local_llog_init(env, obd, lsi->lsi_osd_exp->exp_obd);
817         if (rc)
818                 GOTO(out_llog, rc);
819
820         /* We take an obd ref to insure that we can't get to mgc_cleanup
821          * without calling mgc_fs_cleanup first. */
822         class_incref(obd, "mgc_fs", obd);
823
824         /* We keep the cl_mgc_sem until mgc_fs_cleanup */
825         EXIT;
826 out_llog:
827         if (rc) {
828                 dt_object_put(env, cli->cl_mgc_configs_dir);
829                 cli->cl_mgc_configs_dir = NULL;
830         }
831 out_los:
832         if (rc < 0) {
833                 local_oid_storage_fini(env, cli->cl_mgc_los);
834                 cli->cl_mgc_los = NULL;
835                 mutex_unlock(&cli->cl_mgc_mutex);
836         }
837         return rc;
838 }
839
840 static int mgc_fs_cleanup(const struct lu_env *env, struct obd_device *obd)
841 {
842         struct client_obd       *cli = &obd->u.cli;
843         ENTRY;
844
845         LASSERT(cli->cl_mgc_los != NULL);
846
847         mgc_local_llog_fini(env, obd);
848
849         dt_object_put_nocache(env, cli->cl_mgc_configs_dir);
850         cli->cl_mgc_configs_dir = NULL;
851
852         local_oid_storage_fini(env, cli->cl_mgc_los);
853         cli->cl_mgc_los = NULL;
854
855         class_decref(obd, "mgc_fs", obd);
856         mutex_unlock(&cli->cl_mgc_mutex);
857
858         RETURN(0);
859 }
860
861 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
862 {
863         struct llog_ctxt        *ctxt;
864         int                      rc;
865
866         ENTRY;
867
868         /* setup only remote ctxt, the local disk context is switched per each
869          * filesystem during mgc_fs_setup() */
870         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
871                         &llog_client_ops);
872         if (rc)
873                 RETURN(rc);
874
875         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
876         LASSERT(ctxt);
877
878         llog_initiator_connect(ctxt);
879         llog_ctxt_put(ctxt);
880
881         RETURN(0);
882 }
883
884 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
885 {
886         struct llog_ctxt *ctxt;
887
888         ENTRY;
889
890         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
891         if (ctxt)
892                 llog_cleanup(env, ctxt);
893
894         RETURN(0);
895 }
896
897
898 static atomic_t mgc_count = ATOMIC_INIT(0);
899 static int mgc_precleanup(struct obd_device *obd)
900 {
901         int     rc = 0;
902         int     temp;
903         ENTRY;
904
905         if (atomic_dec_and_test(&mgc_count)) {
906                 LASSERT(rq_state & RQ_RUNNING);
907                 /* stop requeue thread */
908                 temp = RQ_STOP;
909         } else {
910                 /* wakeup requeue thread to clean our cld */
911                 temp = RQ_NOW | RQ_PRECLEANUP;
912         }
913
914         spin_lock(&config_list_lock);
915         rq_state |= temp;
916         spin_unlock(&config_list_lock);
917         wake_up(&rq_waitq);
918
919         if (temp & RQ_STOP)
920                 wait_for_completion(&rq_exit);
921         obd_cleanup_client_import(obd);
922
923         rc = mgc_llog_fini(NULL, obd);
924         if (rc != 0)
925                 CERROR("failed to cleanup llogging subsystems\n");
926
927         RETURN(rc);
928 }
929
930 static int mgc_cleanup(struct obd_device *obd)
931 {
932         int rc;
933         ENTRY;
934
935         /* COMPAT_146 - old config logs may have added profiles we don't
936            know about */
937         if (obd->obd_type->typ_refcnt <= 1)
938                 /* Only for the last mgc */
939                 class_del_profiles();
940
941         lprocfs_obd_cleanup(obd);
942         ptlrpcd_decref();
943
944         rc = client_obd_cleanup(obd);
945         RETURN(rc);
946 }
947
948 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
949 {
950         struct task_struct      *task;
951         int                      rc;
952         ENTRY;
953
954         rc = ptlrpcd_addref();
955         if (rc < 0)
956                 RETURN(rc);
957
958         rc = client_obd_setup(obd, lcfg);
959         if (rc)
960                 GOTO(err_decref, rc);
961
962         rc = mgc_llog_init(NULL, obd);
963         if (rc) {
964                 CERROR("failed to setup llogging subsystems\n");
965                 GOTO(err_cleanup, rc);
966         }
967
968         rc = mgc_tunables_init(obd);
969         if (rc)
970                 GOTO(err_sysfs, rc);
971
972         if (atomic_inc_return(&mgc_count) == 1) {
973                 rq_state = 0;
974                 init_waitqueue_head(&rq_waitq);
975
976                 /* start requeue thread */
977                 task = kthread_run(mgc_requeue_thread, NULL, "ll_cfg_requeue");
978                 if (IS_ERR(task)) {
979                         rc = PTR_ERR(task);
980                         CERROR("%s: cannot start requeue thread: rc = %d; "
981                                "no more log updates\n",
982                                obd->obd_name, rc);
983                         GOTO(err_sysfs, rc);
984                 }
985                 /* rc is the task_struct pointer of mgc_requeue_thread. */
986                 rc = 0;
987                 wait_for_completion(&rq_start);
988         }
989
990         RETURN(rc);
991
992 err_sysfs:
993         lprocfs_obd_cleanup(obd);
994 err_cleanup:
995         client_obd_cleanup(obd);
996 err_decref:
997         ptlrpcd_decref();
998         RETURN(rc);
999 }
1000
1001 /* based on ll_mdc_blocking_ast */
1002 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1003                             void *data, int flag)
1004 {
1005         struct lustre_handle lockh;
1006         struct config_llog_data *cld = (struct config_llog_data *)data;
1007         int rc = 0;
1008         ENTRY;
1009
1010         switch (flag) {
1011         case LDLM_CB_BLOCKING:
1012                 /* mgs wants the lock, give it up... */
1013                 LDLM_DEBUG(lock, "MGC blocking CB");
1014                 ldlm_lock2handle(lock, &lockh);
1015                 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
1016                 break;
1017         case LDLM_CB_CANCELING:
1018                 /* We've given up the lock, prepare ourselves to update. */
1019                 LDLM_DEBUG(lock, "MGC cancel CB");
1020
1021                 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
1022                        PLDLMRES(lock->l_resource),
1023                        (char *)&lock->l_resource->lr_name.name[0]);
1024
1025                 if (!cld) {
1026                         CDEBUG(D_INFO, "missing data, won't requeue\n");
1027                         break;
1028                 }
1029
1030                 /* held at mgc_process_log(). */
1031                 LASSERT(atomic_read(&cld->cld_refcount) > 0);
1032
1033                 lock->l_ast_data = NULL;
1034                 /* Are we done with this log? */
1035                 if (cld->cld_stopping) {
1036                         CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
1037                                 cld->cld_logname);
1038                         config_log_put(cld);
1039                         break;
1040                 }
1041                 /* Make sure not to re-enqueue when the mgc is stopping
1042                    (we get called from client_disconnect_export) */
1043                 if (lock->l_conn_export == NULL ||
1044                     lock->l_conn_export->exp_obd->u.cli.cl_conn_count == 0) {
1045                         CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
1046                                 cld->cld_logname);
1047                         config_log_put(cld);
1048                         break;
1049                 }
1050
1051                 /* Re-enqueue now */
1052                 mgc_requeue_add(cld);
1053                 config_log_put(cld);
1054                 break;
1055         default:
1056                 LBUG();
1057         }
1058
1059         RETURN(rc);
1060 }
1061
1062 /* Not sure where this should go... */
1063 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
1064  * that we can have a chance to try the secondary MGS if any. */
1065 #define  MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
1066                                 + PING_INTERVAL)
1067 #define  MGC_TARGET_REG_LIMIT 10
1068 #define  MGC_TARGET_REG_LIMIT_MAX RECONNECT_DELAY_MAX
1069 #define  MGC_SEND_PARAM_LIMIT 10
1070
1071 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1072 /* Send parameter to MGS*/
1073 static int mgc_set_mgs_param(struct obd_export *exp,
1074                              struct mgs_send_param *msp)
1075 {
1076         struct ptlrpc_request *req;
1077         struct mgs_send_param *req_msp, *rep_msp;
1078         int rc;
1079         ENTRY;
1080
1081         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1082                                         &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
1083                                         MGS_SET_INFO);
1084         if (!req)
1085                 RETURN(-ENOMEM);
1086
1087         req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1088         if (!req_msp) {
1089                 ptlrpc_req_finished(req);
1090                 RETURN(-ENOMEM);
1091         }
1092
1093         memcpy(req_msp, msp, sizeof(*req_msp));
1094         ptlrpc_request_set_replen(req);
1095
1096         /* Limit how long we will wait for the enqueue to complete */
1097         req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
1098         rc = ptlrpc_queue_wait(req);
1099         if (!rc) {
1100                 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
1101                 memcpy(msp, rep_msp, sizeof(*rep_msp));
1102         }
1103
1104         ptlrpc_req_finished(req);
1105
1106         RETURN(rc);
1107 }
1108 #endif
1109
1110 /* Take a config lock so we can get cancel notifications */
1111 static int mgc_enqueue(struct obd_export *exp, enum ldlm_type type,
1112                        union ldlm_policy_data *policy, enum ldlm_mode mode,
1113                        __u64 *flags, ldlm_glimpse_callback glimpse_callback,
1114                        void *data, __u32 lvb_len, void *lvb_swabber,
1115                        struct lustre_handle *lockh)
1116 {
1117         struct config_llog_data *cld = (struct config_llog_data *)data;
1118         struct ldlm_enqueue_info einfo = {
1119                 .ei_type        = type,
1120                 .ei_mode        = mode,
1121                 .ei_cb_bl       = mgc_blocking_ast,
1122                 .ei_cb_cp       = ldlm_completion_ast,
1123                 .ei_cb_gl       = glimpse_callback,
1124         };
1125         struct ptlrpc_request *req;
1126         int short_limit = cld_is_sptlrpc(cld);
1127         int rc;
1128         ENTRY;
1129
1130         CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
1131                cld->cld_resid.name[0]);
1132
1133         /* We need a callback for every lockholder, so don't try to
1134            ldlm_lock_match (see rev 1.1.2.11.2.47) */
1135         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1136                                         &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
1137                                         LDLM_ENQUEUE);
1138         if (req == NULL)
1139                 RETURN(-ENOMEM);
1140
1141         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
1142         ptlrpc_request_set_replen(req);
1143
1144         /* check if this is server or client */
1145         if (cld->cld_cfg.cfg_sb) {
1146                 struct lustre_sb_info *lsi = s2lsi(cld->cld_cfg.cfg_sb);
1147                 if (lsi && IS_SERVER(lsi))
1148                         short_limit = 1;
1149         }
1150         /* Limit how long we will wait for the enqueue to complete */
1151         req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
1152         rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
1153                               NULL, 0, LVB_T_NONE, lockh, 0);
1154         /* A failed enqueue should still call the mgc_blocking_ast,
1155            where it will be requeued if needed ("grant failed"). */
1156         ptlrpc_req_finished(req);
1157         RETURN(rc);
1158 }
1159
1160 static int mgc_cancel(struct obd_export *exp, enum ldlm_mode mode,
1161                       struct lustre_handle *lockh)
1162 {
1163         ENTRY;
1164
1165         ldlm_lock_decref(lockh, mode);
1166
1167         RETURN(0);
1168 }
1169
1170 static void mgc_notify_active(struct obd_device *unused)
1171 {
1172         /* wakeup mgc_requeue_thread to requeue mgc lock */
1173         spin_lock(&config_list_lock);
1174         rq_state |= RQ_NOW;
1175         spin_unlock(&config_list_lock);
1176         wake_up(&rq_waitq);
1177
1178         /* TODO: Help the MGS rebuild nidtbl. -jay */
1179 }
1180
1181 /* Send target_reg message to MGS */
1182 static int mgc_target_register(struct obd_export *exp,
1183                                struct mgs_target_info *mti)
1184 {
1185         struct ptlrpc_request  *req;
1186         struct mgs_target_info *req_mti, *rep_mti;
1187         int                     rc;
1188         ENTRY;
1189
1190         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1191                                         &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
1192                                         MGS_TARGET_REG);
1193         if (req == NULL)
1194                 RETURN(-ENOMEM);
1195
1196         req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
1197         if (!req_mti) {
1198                 ptlrpc_req_finished(req);
1199                 RETURN(-ENOMEM);
1200         }
1201
1202         memcpy(req_mti, mti, sizeof(*req_mti));
1203         ptlrpc_request_set_replen(req);
1204         CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
1205         /* Limit how long we will wait for the enqueue to complete */
1206         req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
1207
1208         /* if the target needs to regenerate the config log in MGS, it's better
1209          * to use some longer limit to let MGC have time to change connection to
1210          * another MGS (or try again with the same MGS) for the target (server)
1211          * will fail and exit if the request expired due to delay limit. */
1212         if (mti->mti_flags & (LDD_F_UPDATE | LDD_F_NEED_INDEX))
1213                 req->rq_delay_limit = MGC_TARGET_REG_LIMIT_MAX;
1214
1215         rc = ptlrpc_queue_wait(req);
1216         if (!rc) {
1217                 rep_mti = req_capsule_server_get(&req->rq_pill,
1218                                                  &RMF_MGS_TARGET_INFO);
1219                 memcpy(mti, rep_mti, sizeof(*rep_mti));
1220                 CDEBUG(D_MGC, "register %s got index = %d\n",
1221                        mti->mti_svname, mti->mti_stripe_index);
1222         }
1223         ptlrpc_req_finished(req);
1224
1225         RETURN(rc);
1226 }
1227
1228 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
1229                               u32 keylen, void *key,
1230                               u32 vallen, void *val,
1231                               struct ptlrpc_request_set *set)
1232 {
1233         int rc = -EINVAL;
1234         ENTRY;
1235
1236         /* Turn off initial_recov after we try all backup servers once */
1237         if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
1238                 struct obd_import *imp = class_exp2cliimp(exp);
1239                 int value;
1240                 if (vallen != sizeof(int))
1241                         RETURN(-EINVAL);
1242                 value = *(int *)val;
1243                 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
1244                        imp->imp_obd->obd_name, value,
1245                        imp->imp_deactive, imp->imp_invalid,
1246                        imp->imp_replayable, imp->imp_obd->obd_replayable,
1247                        ptlrpc_import_state_name(imp->imp_state));
1248                 /* Resurrect the import immediately if
1249                  * 1. we previously got disconnected,
1250                  * 2. value > 1 (at the same node with MGS)
1251                  * */
1252                 if (imp->imp_state == LUSTRE_IMP_DISCON || value > 1)
1253                         ptlrpc_reconnect_import(imp);
1254
1255                 RETURN(0);
1256         }
1257
1258         /* FIXME move this to mgc_process_config */
1259         if (KEY_IS(KEY_REGISTER_TARGET)) {
1260                 struct mgs_target_info *mti;
1261                 if (vallen != sizeof(struct mgs_target_info))
1262                         RETURN(-EINVAL);
1263                 mti = (struct mgs_target_info *)val;
1264                 CDEBUG(D_MGC, "register_target %s %#x\n",
1265                        mti->mti_svname, mti->mti_flags);
1266                 rc =  mgc_target_register(exp, mti);
1267                 RETURN(rc);
1268         }
1269         if (KEY_IS(KEY_SET_FS)) {
1270                 struct super_block *sb = (struct super_block *)val;
1271
1272                 if (vallen != sizeof(struct super_block))
1273                         RETURN(-EINVAL);
1274
1275                 rc = mgc_fs_setup(env, exp->exp_obd, sb);
1276                 RETURN(rc);
1277         }
1278         if (KEY_IS(KEY_CLEAR_FS)) {
1279                 if (vallen != 0)
1280                         RETURN(-EINVAL);
1281                 rc = mgc_fs_cleanup(env, exp->exp_obd);
1282                 RETURN(rc);
1283         }
1284 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 13, 53, 0)
1285         if (KEY_IS(KEY_SET_INFO)) {
1286                 struct mgs_send_param *msp;
1287
1288                 msp = (struct mgs_send_param *)val;
1289                 rc =  mgc_set_mgs_param(exp, msp);
1290                 RETURN(rc);
1291         }
1292 #endif
1293         if (KEY_IS(KEY_MGSSEC)) {
1294                 struct client_obd     *cli = &exp->exp_obd->u.cli;
1295                 struct sptlrpc_flavor  flvr;
1296
1297                 /*
1298                  * empty string means using current flavor, if which haven't
1299                  * been set yet, set it as null.
1300                  *
1301                  * if flavor has been set previously, check the asking flavor
1302                  * must match the existing one.
1303                  */
1304                 if (vallen == 0) {
1305                         if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1306                                 RETURN(0);
1307                         val = "null";
1308                         vallen = 4;
1309                 }
1310
1311                 rc = sptlrpc_parse_flavor(val, &flvr);
1312                 if (rc) {
1313                         CERROR("invalid sptlrpc flavor %s to MGS\n",
1314                                (char *) val);
1315                         RETURN(rc);
1316                 }
1317
1318                 /*
1319                  * caller already hold a mutex
1320                  */
1321                 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1322                         cli->cl_flvr_mgc = flvr;
1323                 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1324                                   sizeof(flvr)) != 0) {
1325                         char    str[20];
1326
1327                         sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1328                                             str, sizeof(str));
1329                         LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but "
1330                                        "currently %s is in use\n",
1331                                        (char *) val, str);
1332                         rc = -EPERM;
1333                 }
1334                 RETURN(rc);
1335         }
1336
1337         RETURN(rc);
1338 }
1339
1340 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1341                         __u32 keylen, void *key, __u32 *vallen, void *val)
1342 {
1343         int rc = -EINVAL;
1344
1345         if (KEY_IS(KEY_CONN_DATA)) {
1346                 struct obd_import *imp = class_exp2cliimp(exp);
1347                 struct obd_connect_data *data = val;
1348
1349                 if (*vallen == sizeof(*data)) {
1350                         *data = imp->imp_connect_data;
1351                         rc = 0;
1352                 }
1353         }
1354
1355         return rc;
1356 }
1357
1358 static int mgc_import_event(struct obd_device *obd,
1359                             struct obd_import *imp,
1360                             enum obd_import_event event)
1361 {
1362         int rc = 0;
1363
1364         LASSERT(imp->imp_obd == obd);
1365         CDEBUG(D_MGC, "import event %#x\n", event);
1366
1367         switch (event) {
1368         case IMP_EVENT_DISCON:
1369                 /* MGC imports should not wait for recovery */
1370                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1371                         ptlrpc_pinger_ir_down();
1372                 break;
1373         case IMP_EVENT_INACTIVE:
1374                 break;
1375         case IMP_EVENT_INVALIDATE: {
1376                 struct ldlm_namespace *ns = obd->obd_namespace;
1377                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1378                 break;
1379         }
1380         case IMP_EVENT_ACTIVE:
1381                 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1382                 /* Clearing obd_no_recov allows us to continue pinging */
1383                 obd->obd_no_recov = 0;
1384                 mgc_notify_active(obd);
1385                 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1386                         ptlrpc_pinger_ir_up();
1387                 break;
1388         case IMP_EVENT_OCD:
1389                 break;
1390         case IMP_EVENT_DEACTIVATE:
1391         case IMP_EVENT_ACTIVATE:
1392                 break;
1393         default:
1394                 CERROR("Unknown import event %#x\n", event);
1395                 LBUG();
1396         }
1397         RETURN(rc);
1398 }
1399
1400 enum {
1401         CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_SHIFT),
1402         CONFIG_READ_NRPAGES      = 4
1403 };
1404
1405 static int mgc_apply_recover_logs(struct obd_device *mgc,
1406                                   struct config_llog_data *cld,
1407                                   __u64 max_version,
1408                                   void *data, int datalen, bool mne_swab)
1409 {
1410         struct config_llog_instance *cfg = &cld->cld_cfg;
1411         struct lustre_sb_info       *lsi = s2lsi(cfg->cfg_sb);
1412         struct mgs_nidtbl_entry *entry;
1413         struct lustre_cfg       *lcfg;
1414         struct lustre_cfg_bufs   bufs;
1415         u64   prev_version = 0;
1416         char *inst;
1417         char *buf;
1418         int   bufsz;
1419         int   pos;
1420         int   rc  = 0;
1421         int   off = 0;
1422         ENTRY;
1423
1424         LASSERT(cfg->cfg_instance != NULL);
1425         LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1426
1427         OBD_ALLOC(inst, PAGE_SIZE);
1428         if (inst == NULL)
1429                 RETURN(-ENOMEM);
1430
1431         if (!IS_SERVER(lsi)) {
1432                 pos = snprintf(inst, PAGE_SIZE, "%p", cfg->cfg_instance);
1433                 if (pos >= PAGE_SIZE) {
1434                         OBD_FREE(inst, PAGE_SIZE);
1435                         return -E2BIG;
1436                 }
1437         } else {
1438                 LASSERT(IS_MDT(lsi));
1439                 rc = server_name2svname(lsi->lsi_svname, inst, NULL,
1440                                         PAGE_SIZE);
1441                 if (rc) {
1442                         OBD_FREE(inst, PAGE_SIZE);
1443                         RETURN(-EINVAL);
1444                 }
1445                 pos = strlen(inst);
1446         }
1447
1448         ++pos;
1449         buf   = inst + pos;
1450         bufsz = PAGE_SIZE - pos;
1451
1452         while (datalen > 0) {
1453                 int   entry_len = sizeof(*entry);
1454                 int   is_ost, i;
1455                 struct obd_device *obd;
1456                 char *obdname;
1457                 char *cname;
1458                 char *params;
1459                 char *uuid;
1460
1461                 rc = -EINVAL;
1462                 if (datalen < sizeof(*entry))
1463                         break;
1464
1465                 entry = (typeof(entry))(data + off);
1466
1467                 /* sanity check */
1468                 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1469                         break;
1470                 if (entry->mne_nid_count == 0) /* at least one nid entry */
1471                         break;
1472                 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1473                         break;
1474
1475                 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1476                 if (datalen < entry_len) /* must have entry_len at least */
1477                         break;
1478
1479                 /* Keep this swab for normal mixed endian handling. LU-1644 */
1480                 if (mne_swab)
1481                         lustre_swab_mgs_nidtbl_entry(entry);
1482                 if (entry->mne_length > PAGE_SIZE) {
1483                         CERROR("MNE too large (%u)\n", entry->mne_length);
1484                         break;
1485                 }
1486
1487                 if (entry->mne_length < entry_len)
1488                         break;
1489
1490                 off     += entry->mne_length;
1491                 datalen -= entry->mne_length;
1492                 if (datalen < 0)
1493                         break;
1494
1495                 if (entry->mne_version > max_version) {
1496                         CERROR("entry index(%lld) is over max_index(%lld)\n",
1497                                entry->mne_version, max_version);
1498                         break;
1499                 }
1500
1501                 if (prev_version >= entry->mne_version) {
1502                         CERROR("index unsorted, prev %lld, now %lld\n",
1503                                prev_version, entry->mne_version);
1504                         break;
1505                 }
1506                 prev_version = entry->mne_version;
1507
1508                 /*
1509                  * Write a string with format "nid::instance" to
1510                  * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1511                  */
1512
1513                 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1514                 memset(buf, 0, bufsz);
1515                 obdname = buf;
1516                 pos = 0;
1517
1518                 /* lustre-OST0001-osc-<instance #> */
1519                 strcpy(obdname, cld->cld_logname);
1520                 cname = strrchr(obdname, '-');
1521                 if (cname == NULL) {
1522                         CERROR("mgc %s: invalid logname %s\n",
1523                                mgc->obd_name, obdname);
1524                         break;
1525                 }
1526
1527                 pos = cname - obdname;
1528                 obdname[pos] = 0;
1529                 pos += sprintf(obdname + pos, "-%s%04x",
1530                                   is_ost ? "OST" : "MDT", entry->mne_index);
1531
1532                 cname = is_ost ? "osc" : "mdc",
1533                 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1534                 lustre_cfg_bufs_reset(&bufs, obdname);
1535
1536                 /* find the obd by obdname */
1537                 obd = class_name2obd(obdname);
1538                 if (obd == NULL) {
1539                         CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1540                                mgc->obd_name, obdname);
1541                         rc = 0;
1542                         /* this is a safe race, when the ost is starting up...*/
1543                         continue;
1544                 }
1545
1546                 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1547                 ++pos;
1548                 params = buf + pos;
1549                 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1550                 uuid = buf + pos;
1551
1552                 down_read(&obd->u.cli.cl_sem);
1553                 if (obd->u.cli.cl_import == NULL) {
1554                         /* client does not connect to the OST yet */
1555                         up_read(&obd->u.cli.cl_sem);
1556                         rc = 0;
1557                         continue;
1558                 }
1559
1560                 /* iterate all nids to find one */
1561                 /* find uuid by nid */
1562                 rc = -ENOENT;
1563                 for (i = 0; i < entry->mne_nid_count; i++) {
1564                         rc = client_import_find_conn(obd->u.cli.cl_import,
1565                                                      entry->u.nids[i],
1566                                                      (struct obd_uuid *)uuid);
1567                         if (rc == 0)
1568                                 break;
1569                 }
1570
1571                 up_read(&obd->u.cli.cl_sem);
1572                 if (rc < 0) {
1573                         CERROR("mgc: cannot find uuid by nid %s\n",
1574                                libcfs_nid2str(entry->u.nids[0]));
1575                         break;
1576                 }
1577
1578                 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1579                        uuid, libcfs_nid2str(entry->u.nids[0]));
1580
1581                 pos += strlen(uuid);
1582                 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1583                 LASSERT(pos < bufsz);
1584
1585                 lustre_cfg_bufs_set_string(&bufs, 1, params);
1586
1587                 OBD_ALLOC(lcfg, lustre_cfg_len(bufs.lcfg_bufcount,
1588                                                bufs.lcfg_buflen));
1589                 if (!lcfg) {
1590                         rc = -ENOMEM;
1591                         break;
1592                 }
1593                 lustre_cfg_init(lcfg, LCFG_PARAM, &bufs);
1594
1595                 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1596                        prev_version, max_version, obdname, params);
1597
1598                 rc = class_process_config(lcfg);
1599                 OBD_FREE(lcfg, lustre_cfg_len(lcfg->lcfg_bufcount,
1600                                               lcfg->lcfg_buflens));
1601                 if (rc)
1602                         CDEBUG(D_INFO, "process config for %s error %d\n",
1603                                obdname, rc);
1604
1605                 /* continue, even one with error */
1606         }
1607
1608         OBD_FREE(inst, PAGE_SIZE);
1609         RETURN(rc);
1610 }
1611
1612 /**
1613  * This function is called if this client was notified for target restarting
1614  * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery or
1615  * nodemap logs.
1616  */
1617 static int mgc_process_recover_nodemap_log(struct obd_device *obd,
1618                                            struct config_llog_data *cld)
1619 {
1620         struct ptlrpc_connection *mgc_conn;
1621         struct ptlrpc_request *req = NULL;
1622         struct config_llog_instance *cfg = &cld->cld_cfg;
1623         struct mgs_config_body *body;
1624         struct mgs_config_res *res;
1625         struct nodemap_config *new_config = NULL;
1626         struct lu_nodemap *recent_nodemap = NULL;
1627         struct ptlrpc_bulk_desc *desc;
1628         struct page **pages = NULL;
1629         __u64 config_read_offset = 0;
1630         __u8 nodemap_cur_pass = 0;
1631         int nrpages = 0;
1632         bool eof = true;
1633         bool mne_swab = false;
1634         int i;
1635         int ealen;
1636         int rc;
1637         ENTRY;
1638
1639         mgc_conn = class_exp2cliimp(cld->cld_mgcexp)->imp_connection;
1640
1641         /* don't need to get local config */
1642         if (cld_is_nodemap(cld) &&
1643             (LNET_NETTYP(LNET_NIDNET(mgc_conn->c_peer.nid)) == LOLND))
1644                 GOTO(out, rc = 0);
1645
1646         /* allocate buffer for bulk transfer.
1647          * if this is the first time for this mgs to read logs,
1648          * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1649          * once; otherwise, it only reads increment of logs, this should be
1650          * small and CONFIG_READ_NRPAGES will be used.
1651          */
1652         nrpages = CONFIG_READ_NRPAGES;
1653         if (cfg->cfg_last_idx == 0 || cld_is_nodemap(cld))
1654                 nrpages = CONFIG_READ_NRPAGES_INIT;
1655
1656         OBD_ALLOC(pages, sizeof(*pages) * nrpages);
1657         if (pages == NULL)
1658                 GOTO(out, rc = -ENOMEM);
1659
1660         for (i = 0; i < nrpages; i++) {
1661                 pages[i] = alloc_page(GFP_KERNEL);
1662                 if (pages[i] == NULL)
1663                         GOTO(out, rc = -ENOMEM);
1664         }
1665
1666 again:
1667 #ifdef HAVE_SERVER_SUPPORT
1668         if (cld_is_nodemap(cld) && config_read_offset == 0) {
1669                 new_config = nodemap_config_alloc();
1670                 if (IS_ERR(new_config)) {
1671                         rc = PTR_ERR(new_config);
1672                         new_config = NULL;
1673                         GOTO(out, rc);
1674                 }
1675         }
1676 #endif
1677         LASSERT(cld_is_recover(cld) || cld_is_nodemap(cld));
1678         LASSERT(mutex_is_locked(&cld->cld_lock));
1679         req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1680                                    &RQF_MGS_CONFIG_READ);
1681         if (req == NULL)
1682                 GOTO(out, rc = -ENOMEM);
1683
1684         rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1685         if (rc)
1686                 GOTO(out, rc);
1687
1688         /* pack request */
1689         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1690         LASSERT(body != NULL);
1691         LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1692         if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1693             >= sizeof(body->mcb_name))
1694                 GOTO(out, rc = -E2BIG);
1695         if (cld_is_nodemap(cld))
1696                 body->mcb_offset = config_read_offset;
1697         else
1698                 body->mcb_offset = cfg->cfg_last_idx + 1;
1699         body->mcb_type   = cld->cld_type;
1700         body->mcb_bits   = PAGE_SHIFT;
1701         body->mcb_units  = nrpages;
1702         body->mcb_nm_cur_pass = nodemap_cur_pass;
1703
1704         /* allocate bulk transfer descriptor */
1705         desc = ptlrpc_prep_bulk_imp(req, nrpages, 1,
1706                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1707                                     MGS_BULK_PORTAL,
1708                                     &ptlrpc_bulk_kiov_pin_ops);
1709         if (desc == NULL)
1710                 GOTO(out, rc = -ENOMEM);
1711
1712         for (i = 0; i < nrpages; i++)
1713                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1714                                                  PAGE_SIZE);
1715
1716         ptlrpc_request_set_replen(req);
1717         rc = ptlrpc_queue_wait(req);
1718         if (rc)
1719                 GOTO(out, rc);
1720
1721         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1722         if (!res)
1723                 GOTO(out, rc = -EPROTO);
1724
1725         if (cld_is_nodemap(cld)) {
1726                 config_read_offset = res->mcr_offset;
1727                 eof = config_read_offset == II_END_OFF;
1728                 nodemap_cur_pass = res->mcr_nm_cur_pass;
1729         } else {
1730                 if (res->mcr_size < res->mcr_offset)
1731                         GOTO(out, rc = -EINVAL);
1732
1733                 /* always update the index even though it might have errors with
1734                  * handling the recover logs
1735                  */
1736                 cfg->cfg_last_idx = res->mcr_offset;
1737                 eof = res->mcr_offset == res->mcr_size;
1738
1739                 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1740                        res->mcr_offset, eof == false);
1741         }
1742
1743         ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1744         if (ealen < 0)
1745                 GOTO(out, rc = ealen);
1746
1747         if (ealen > nrpages << PAGE_SHIFT)
1748                 GOTO(out, rc = -EINVAL);
1749
1750         if (ealen == 0) { /* no logs transferred */
1751 #ifdef HAVE_SERVER_SUPPORT
1752                 /* config changed since first read RPC */
1753                 if (cld_is_nodemap(cld) && config_read_offset == 0) {
1754                         CDEBUG(D_INFO, "nodemap config changed in transit, retrying\n");
1755                         GOTO(out, rc = -EAGAIN);
1756                 }
1757 #endif
1758                 if (!eof)
1759                         rc = -EINVAL;
1760                 GOTO(out, rc);
1761         }
1762
1763         mne_swab = ptlrpc_rep_need_swab(req);
1764
1765         /* When a nodemap config is received, we build a new nodemap config,
1766          * with new nodemap structs. We keep track of the most recently added
1767          * nodemap since the config is read ordered by nodemap_id, and so it
1768          * is likely that the next record will be related. Because access to
1769          * the nodemaps is single threaded until the nodemap_config is active,
1770          * we don't need to reference count with recent_nodemap, though
1771          * recent_nodemap should be set to NULL when the nodemap_config
1772          * is either destroyed or set active.
1773          */
1774         for (i = 0; i < nrpages && ealen > 0; i++) {
1775                 int rc2;
1776                 union lu_page   *ptr;
1777
1778                 ptr = kmap(pages[i]);
1779                 if (cld_is_nodemap(cld))
1780                         rc2 = nodemap_process_idx_pages(new_config, ptr,
1781                                                        &recent_nodemap);
1782                 else
1783                         rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset,
1784                                                      ptr,
1785                                                      min_t(int, ealen,
1786                                                            PAGE_SIZE),
1787                                                      mne_swab);
1788                 kunmap(pages[i]);
1789                 if (rc2 < 0) {
1790                         CWARN("%s: error processing %s log %s: rc = %d\n",
1791                               obd->obd_name,
1792                               cld_is_nodemap(cld) ? "nodemap" : "recovery",
1793                               cld->cld_logname,
1794                               rc2);
1795                         GOTO(out, rc = rc2);
1796                 }
1797
1798                 ealen -= PAGE_SIZE;
1799         }
1800
1801 out:
1802         if (req) {
1803                 ptlrpc_req_finished(req);
1804                 req = NULL;
1805         }
1806
1807         if (rc == 0 && !eof)
1808                 goto again;
1809
1810 #ifdef HAVE_SERVER_SUPPORT
1811         if (new_config != NULL) {
1812                 /* recent_nodemap cannot be used after set_active/dealloc */
1813                 if (rc == 0)
1814                         nodemap_config_set_active_mgc(new_config);
1815                 else
1816                         nodemap_config_dealloc(new_config);
1817         }
1818 #endif
1819
1820         if (pages) {
1821                 for (i = 0; i < nrpages; i++) {
1822                         if (pages[i] == NULL)
1823                                 break;
1824                         __free_page(pages[i]);
1825                 }
1826                 OBD_FREE(pages, sizeof(*pages) * nrpages);
1827         }
1828         return rc;
1829 }
1830
1831 static int mgc_barrier_glimpse_ast(struct ldlm_lock *lock, void *data)
1832 {
1833         struct config_llog_data *cld = lock->l_ast_data;
1834         int rc;
1835         ENTRY;
1836
1837         if (cld->cld_stopping)
1838                 RETURN(-ENODEV);
1839
1840         rc = barrier_handler(s2lsi(cld->cld_cfg.cfg_sb)->lsi_dt_dev,
1841                              (struct ptlrpc_request *)data);
1842
1843         RETURN(rc);
1844 }
1845
1846 /* Copy a remote log locally */
1847 static int mgc_llog_local_copy(const struct lu_env *env,
1848                                struct obd_device *obd,
1849                                struct llog_ctxt *rctxt,
1850                                struct llog_ctxt *lctxt, char *logname)
1851 {
1852         char    *temp_log;
1853         int      rc;
1854
1855         ENTRY;
1856
1857         /*
1858          * - copy it to backup using llog_backup()
1859          * - copy remote llog to logname using llog_backup()
1860          * - if failed then move bakup to logname again
1861          */
1862
1863         OBD_ALLOC(temp_log, strlen(logname) + 2);
1864         if (!temp_log)
1865                 RETURN(-ENOMEM);
1866         sprintf(temp_log, "%sT", logname);
1867
1868         /* make a copy of local llog at first */
1869         rc = llog_backup(env, obd, lctxt, lctxt, logname, temp_log);
1870         if (rc < 0 && rc != -ENOENT)
1871                 GOTO(out, rc);
1872         /* copy remote llog to the local copy */
1873         rc = llog_backup(env, obd, rctxt, lctxt, logname, logname);
1874         if (rc == -ENOENT) {
1875                 /* no remote llog, delete local one too */
1876                 llog_erase(env, lctxt, NULL, logname);
1877         } else if (rc < 0) {
1878                 /* error during backup, get local one back from the copy */
1879                 llog_backup(env, obd, lctxt, lctxt, temp_log, logname);
1880 out:
1881                 CERROR("%s: failed to copy remote log %s: rc = %d\n",
1882                        obd->obd_name, logname, rc);
1883         }
1884         llog_erase(env, lctxt, NULL, temp_log);
1885         OBD_FREE(temp_log, strlen(logname) + 2);
1886         return rc;
1887 }
1888
1889 /* local_only means it cannot get remote llogs */
1890 static int mgc_process_cfg_log(struct obd_device *mgc,
1891                                struct config_llog_data *cld, int local_only)
1892 {
1893         struct llog_ctxt        *ctxt, *lctxt = NULL;
1894         struct client_obd       *cli = &mgc->u.cli;
1895         struct lustre_sb_info   *lsi = NULL;
1896         int                      rc = 0;
1897         struct lu_env           *env;
1898
1899         ENTRY;
1900
1901         LASSERT(cld);
1902         LASSERT(mutex_is_locked(&cld->cld_lock));
1903
1904         if (cld->cld_cfg.cfg_sb)
1905                 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1906
1907         OBD_ALLOC_PTR(env);
1908         if (env == NULL)
1909                 RETURN(-ENOMEM);
1910
1911         rc = lu_env_init(env, LCT_MG_THREAD);
1912         if (rc)
1913                 GOTO(out_free, rc);
1914
1915         ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1916         LASSERT(ctxt);
1917
1918         lctxt = llog_get_context(mgc, LLOG_CONFIG_ORIG_CTXT);
1919
1920         /* Copy the setup log locally if we can. Don't mess around if we're
1921          * running an MGS though (logs are already local). */
1922         if (lctxt && lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1923             cli->cl_mgc_configs_dir != NULL &&
1924             lu2dt_dev(cli->cl_mgc_configs_dir->do_lu.lo_dev) ==
1925             lsi->lsi_dt_dev) {
1926                 if (!local_only && !lsi->lsi_dt_dev->dd_rdonly)
1927                         /* Only try to copy log if we have the lock. */
1928                         rc = mgc_llog_local_copy(env, mgc, ctxt, lctxt,
1929                                                  cld->cld_logname);
1930                 if (local_only || rc) {
1931                         if (strcmp(cld->cld_logname, PARAMS_FILENAME) != 0 &&
1932                             llog_is_empty(env, lctxt, cld->cld_logname)) {
1933                                 LCONSOLE_ERROR_MSG(0x13a, "Failed to get MGS "
1934                                                    "log %s and no local copy."
1935                                                    "\n", cld->cld_logname);
1936                                 GOTO(out_pop, rc = -ENOENT);
1937                         }
1938                         CDEBUG(D_MGC, "Failed to get MGS log %s, using local "
1939                                "copy for now, will try to update later.\n",
1940                                cld->cld_logname);
1941                         rc = 0;
1942                 }
1943                 /* Now, whether we copied or not, start using the local llog.
1944                  * If we failed to copy, we'll start using whatever the old
1945                  * log has. */
1946                 llog_ctxt_put(ctxt);
1947                 ctxt = lctxt;
1948                 lctxt = NULL;
1949         } else {
1950                 if (local_only) /* no local log at client side */
1951                         GOTO(out_pop, rc = -EIO);
1952         }
1953
1954         rc = -EAGAIN;
1955         if (lsi && IS_SERVER(lsi) && !IS_MGS(lsi) &&
1956             lsi->lsi_dt_dev->dd_rdonly) {
1957                 struct llog_ctxt *rctxt;
1958
1959                 /* Under readonly mode, we may have no local copy or local
1960                  * copy is incomplete, so try to use remote llog firstly. */
1961                 rctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1962                 LASSERT(rctxt);
1963
1964                 rc = class_config_parse_llog(env, rctxt, cld->cld_logname,
1965                                              &cld->cld_cfg);
1966                 llog_ctxt_put(rctxt);
1967         }
1968
1969         if (rc && rc != -ENOENT)
1970                 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1971                                              &cld->cld_cfg);
1972
1973         /*
1974          * update settings on existing OBDs. doing it inside
1975          * of llog_process_lock so no device is attaching/detaching
1976          * in parallel.
1977          * the logname must be <fsname>-sptlrpc
1978          */
1979         if (rc == 0 && cld_is_sptlrpc(cld))
1980                 class_notify_sptlrpc_conf(cld->cld_logname,
1981                                           strlen(cld->cld_logname) -
1982                                           strlen("-sptlrpc"));
1983         EXIT;
1984
1985 out_pop:
1986         __llog_ctxt_put(env, ctxt);
1987         if (lctxt)
1988                 __llog_ctxt_put(env, lctxt);
1989
1990         lu_env_fini(env);
1991 out_free:
1992         OBD_FREE_PTR(env);
1993         return rc;
1994 }
1995
1996 static bool mgc_import_in_recovery(struct obd_import *imp)
1997 {
1998         bool in_recovery = true;
1999
2000         spin_lock(&imp->imp_lock);
2001         if (imp->imp_state == LUSTRE_IMP_FULL ||
2002             imp->imp_state == LUSTRE_IMP_CLOSED)
2003                 in_recovery = false;
2004         spin_unlock(&imp->imp_lock);
2005
2006         return in_recovery;
2007 }
2008
2009 /**
2010  * Get a configuration log from the MGS and process it.
2011  *
2012  * This function is called for both clients and servers to process the
2013  * configuration log from the MGS.  The MGC enqueues a DLM lock on the
2014  * log from the MGS, and if the lock gets revoked the MGC will be notified
2015  * by the lock cancellation callback that the config log has changed,
2016  * and will enqueue another MGS lock on it, and then continue processing
2017  * the new additions to the end of the log.
2018  *
2019  * Since the MGC import is not replayable, if the import is being evicted
2020  * (rcl == -ESHUTDOWN, \see ptlrpc_import_delay_req()), retry to process
2021  * the log until recovery is finished or the import is closed.
2022  *
2023  * Make a local copy of the log before parsing it if appropriate (non-MGS
2024  * server) so that the server can start even when the MGS is down.
2025  *
2026  * There shouldn't be multiple processes running process_log at once --
2027  * sounds like badness.  It actually might be fine, as long as they're not
2028  * trying to update from the same log simultaneously, in which case we
2029  * should use a per-log semaphore instead of cld_lock.
2030  *
2031  * \param[in] mgc       MGC device by which to fetch the configuration log
2032  * \param[in] cld       log processing state (stored in lock callback data)
2033  *
2034  * \retval              0 on success
2035  * \retval              negative errno on failure
2036  */
2037 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
2038 {
2039         struct lustre_handle lockh = { 0 };
2040         __u64 flags = LDLM_FL_NO_LRU;
2041         int rc = 0, rcl;
2042         bool retry = false;
2043         ENTRY;
2044
2045         LASSERT(cld != NULL);
2046
2047         /* I don't want multiple processes running process_log at once --
2048            sounds like badness.  It actually might be fine, as long as
2049            we're not trying to update from the same log
2050            simultaneously (in which case we should use a per-log sem.) */
2051 restart:
2052         mutex_lock(&cld->cld_lock);
2053         if (cld->cld_stopping) {
2054                 mutex_unlock(&cld->cld_lock);
2055                 RETURN(0);
2056         }
2057
2058         OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
2059
2060         CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
2061                cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
2062
2063         /* Get the cfg lock on the llog */
2064         rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, LDLM_PLAIN, NULL,
2065                           LCK_CR, &flags,
2066                           cld_is_barrier(cld) ? mgc_barrier_glimpse_ast : NULL,
2067                           cld, 0, NULL, &lockh);
2068         if (rcl == 0) {
2069                 /* Get the cld, it will be released in mgc_blocking_ast. */
2070                 config_log_get(cld);
2071                 rc = ldlm_lock_set_data(&lockh, (void *)cld);
2072                 LASSERT(rc == 0);
2073         } else {
2074                 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
2075
2076                 if (rcl == -ESHUTDOWN &&
2077                     atomic_read(&mgc->u.cli.cl_mgc_refcount) > 0 && !retry) {
2078                         struct obd_import *imp;
2079                         struct l_wait_info lwi;
2080                         int secs = cfs_time_seconds(obd_timeout);
2081
2082                         mutex_unlock(&cld->cld_lock);
2083                         imp = class_exp2cliimp(mgc->u.cli.cl_mgc_mgsexp);
2084
2085                         /* Let's force the pinger, and wait the import to be
2086                          * connected, note: since mgc import is non-replayable,
2087                          * and even the import state is disconnected, it does
2088                          * not mean the "recovery" is stopped, so we will keep
2089                          * waitting until timeout or the import state is
2090                          * FULL or closed */
2091                         ptlrpc_pinger_force(imp);
2092
2093                         lwi = LWI_TIMEOUT(secs, NULL, NULL);
2094                         l_wait_event(imp->imp_recovery_waitq,
2095                                      !mgc_import_in_recovery(imp), &lwi);
2096
2097                         if (imp->imp_state == LUSTRE_IMP_FULL) {
2098                                 retry = true;
2099                                 goto restart;
2100                         } else {
2101                                 mutex_lock(&cld->cld_lock);
2102                                 spin_lock(&config_list_lock);
2103                                 cld->cld_lostlock = 1;
2104                                 spin_unlock(&config_list_lock);
2105                         }
2106                 } else {
2107                         /* mark cld_lostlock so that it will requeue
2108                          * after MGC becomes available. */
2109                         spin_lock(&config_list_lock);
2110                         cld->cld_lostlock = 1;
2111                         spin_unlock(&config_list_lock);
2112                 }
2113         }
2114
2115         if (cld_is_recover(cld) || cld_is_nodemap(cld)) {
2116                 if (!rcl)
2117                         rc = mgc_process_recover_nodemap_log(mgc, cld);
2118                 else if (cld_is_nodemap(cld))
2119                         rc = rcl;
2120
2121                 if (cld_is_recover(cld) && rc) {
2122                         if (!rcl) {
2123                                 CERROR("%s: recover log %s failed, not fatal: rc = %d\n",
2124                                        mgc->obd_name, cld->cld_logname, rc);
2125                                 spin_lock(&config_list_lock);
2126                                 cld->cld_lostlock = 1;
2127                                 spin_unlock(&config_list_lock);
2128                         }
2129                         rc = 0; /* this is not a fatal error for recover log */
2130                 }
2131         } else if (!cld_is_barrier(cld)) {
2132                 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
2133         }
2134
2135         CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
2136                mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
2137
2138         mutex_unlock(&cld->cld_lock);
2139
2140         /* Now drop the lock so MGS can revoke it */
2141         if (!rcl) {
2142                 rcl = mgc_cancel(mgc->u.cli.cl_mgc_mgsexp, LCK_CR, &lockh);
2143                 if (rcl)
2144                         CERROR("Can't drop cfg lock: %d\n", rcl);
2145         }
2146
2147         /* requeue nodemap lock immediately if transfer was interrupted */
2148         if (cld_is_nodemap(cld) && rc == -EAGAIN) {
2149                 mgc_requeue_add(cld);
2150                 rc = 0;
2151         }
2152
2153         RETURN(rc);
2154 }
2155
2156
2157 /** Called from lustre_process_log.
2158  * LCFG_LOG_START gets the config log from the MGS, processes it to start
2159  * any services, and adds it to the list logs to watch (follow).
2160  */
2161 static int mgc_process_config(struct obd_device *obd, size_t len, void *buf)
2162 {
2163         struct lustre_cfg *lcfg = buf;
2164         struct config_llog_instance *cfg = NULL;
2165         char *logname;
2166         int rc = 0;
2167         ENTRY;
2168
2169         switch(lcfg->lcfg_command) {
2170         case LCFG_LOV_ADD_OBD: {
2171                 /* Overloading this cfg command: register a new target */
2172                 struct mgs_target_info *mti;
2173
2174                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
2175                     sizeof(struct mgs_target_info))
2176                         GOTO(out, rc = -EINVAL);
2177
2178                 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
2179                 CDEBUG(D_MGC, "add_target %s %#x\n",
2180                        mti->mti_svname, mti->mti_flags);
2181                 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
2182                 break;
2183         }
2184         case LCFG_LOV_DEL_OBD:
2185                 /* Unregister has no meaning at the moment. */
2186                 CERROR("lov_del_obd unimplemented\n");
2187                 rc = -ENOSYS;
2188                 break;
2189         case LCFG_SPTLRPC_CONF: {
2190                 rc = sptlrpc_process_config(lcfg);
2191                 break;
2192         }
2193         case LCFG_LOG_START: {
2194                 struct config_llog_data *cld;
2195                 struct super_block *sb;
2196
2197                 logname = lustre_cfg_string(lcfg, 1);
2198                 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
2199                 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
2200
2201                 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
2202                        cfg->cfg_last_idx);
2203
2204                 /* We're only called through here on the initial mount */
2205                 cld = config_log_add(obd, logname, cfg, sb);
2206                 if (IS_ERR(cld)) {
2207                         rc = PTR_ERR(cld);
2208                         break;
2209                 }
2210
2211                 rc = mgc_process_log(obd, cld);
2212                 if (rc == 0 && cld->cld_recover != NULL) {
2213                         if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
2214                                          imp_connect_data, IMP_RECOV)) {
2215                                 rc = mgc_process_log(obd, cld->cld_recover);
2216                         } else {
2217                                 struct config_llog_data *cir;
2218
2219                                 mutex_lock(&cld->cld_lock);
2220                                 cir = cld->cld_recover;
2221                                 cld->cld_recover = NULL;
2222                                 mutex_unlock(&cld->cld_lock);
2223                                 config_log_put(cir);
2224                         }
2225
2226                         if (rc)
2227                                 CERROR("Cannot process recover llog %d\n", rc);
2228                 }
2229
2230                 if (rc == 0 && cld->cld_params != NULL) {
2231                         rc = mgc_process_log(obd, cld->cld_params);
2232                         if (rc == -ENOENT) {
2233                                 CDEBUG(D_MGC, "There is no params "
2234                                               "config file yet\n");
2235                                 rc = 0;
2236                         }
2237                         /* params log is optional */
2238                         if (rc)
2239                                 CERROR("%s: can't process params llog: rc = %d\n",
2240                                        obd->obd_name, rc);
2241                 }
2242
2243                 break;
2244         }
2245         case LCFG_LOG_END: {
2246                 logname = lustre_cfg_string(lcfg, 1);
2247
2248                 if (lcfg->lcfg_bufcount >= 2)
2249                         cfg = (struct config_llog_instance *)lustre_cfg_buf(
2250                                 lcfg, 2);
2251                 rc = config_log_end(logname, cfg);
2252                 break;
2253         }
2254         default: {
2255                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
2256                 GOTO(out, rc = -EINVAL);
2257
2258         }
2259         }
2260 out:
2261         RETURN(rc);
2262 }
2263
2264 static struct obd_ops mgc_obd_ops = {
2265         .o_owner        = THIS_MODULE,
2266         .o_setup        = mgc_setup,
2267         .o_precleanup   = mgc_precleanup,
2268         .o_cleanup      = mgc_cleanup,
2269         .o_add_conn     = client_import_add_conn,
2270         .o_del_conn     = client_import_del_conn,
2271         .o_connect      = client_connect_import,
2272         .o_disconnect   = client_disconnect_export,
2273         .o_set_info_async = mgc_set_info_async,
2274         .o_get_info       = mgc_get_info,
2275         .o_import_event = mgc_import_event,
2276         .o_process_config = mgc_process_config,
2277 };
2278
2279 static int __init mgc_init(void)
2280 {
2281         return class_register_type(&mgc_obd_ops, NULL, false, NULL,
2282                                    LUSTRE_MGC_NAME, NULL);
2283 }
2284
2285 static void __exit mgc_exit(void)
2286 {
2287         class_unregister_type(LUSTRE_MGC_NAME);
2288 }
2289
2290 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2291 MODULE_DESCRIPTION("Lustre Management Client");
2292 MODULE_VERSION(LUSTRE_VERSION_STRING);
2293 MODULE_LICENSE("GPL");
2294
2295 module_init(mgc_init);
2296 module_exit(mgc_exit);