1 /*
2 * GPL HEADER START
3 *
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
19 *
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
22 * have any questions.
23 *
24 * GPL HEADER END
25 */
26 /*
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
29 *
30 * Copyright (c) 2011, 2012, Intel Corporation.
31 */
32 /*
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
35 *
36 * lustre/mgc/mgc_request.c
37 *
38 * Author: Nathan Rutman <nathan@clusterfs.com>
39 */
40
41 #define DEBUG_SUBSYSTEM S_MGC
42 #define D_MGC D_CONFIG /*|D_WARNING*/
43
44 #include <linux/module.h>
45 #include "../include/obd_class.h"
46 #include "../include/lustre_dlm.h"
47 #include "../include/lprocfs_status.h"
48 #include "../include/lustre_log.h"
49 #include "../include/lustre_disk.h"
50
51 #include "mgc_internal.h"
52
mgc_name2resid(char * name,int len,struct ldlm_res_id * res_id,int type)53 static int mgc_name2resid(char *name, int len, struct ldlm_res_id *res_id,
54 int type)
55 {
56 __u64 resname = 0;
57
58 if (len > sizeof(resname)) {
59 CERROR("name too long: %s\n", name);
60 return -EINVAL;
61 }
62 if (len <= 0) {
63 CERROR("missing name: %s\n", name);
64 return -EINVAL;
65 }
66 memcpy(&resname, name, len);
67
68 /* Always use the same endianness for the resid */
69 memset(res_id, 0, sizeof(*res_id));
70 res_id->name[0] = cpu_to_le64(resname);
71 /* XXX: unfortunately, sptlprc and config llog share one lock */
72 switch (type) {
73 case CONFIG_T_CONFIG:
74 case CONFIG_T_SPTLRPC:
75 resname = 0;
76 break;
77 case CONFIG_T_RECOVER:
78 case CONFIG_T_PARAMS:
79 resname = type;
80 break;
81 default:
82 LBUG();
83 }
84 res_id->name[1] = cpu_to_le64(resname);
85 CDEBUG(D_MGC, "log %s to resid %#llx/%#llx (%.8s)\n", name,
86 res_id->name[0], res_id->name[1], (char *)&res_id->name[0]);
87 return 0;
88 }
89
mgc_fsname2resid(char * fsname,struct ldlm_res_id * res_id,int type)90 int mgc_fsname2resid(char *fsname, struct ldlm_res_id *res_id, int type)
91 {
92 /* fsname is at most 8 chars long, maybe contain "-".
93 * e.g. "lustre", "SUN-000" */
94 return mgc_name2resid(fsname, strlen(fsname), res_id, type);
95 }
96 EXPORT_SYMBOL(mgc_fsname2resid);
97
mgc_logname2resid(char * logname,struct ldlm_res_id * res_id,int type)98 static int mgc_logname2resid(char *logname, struct ldlm_res_id *res_id, int type)
99 {
100 char *name_end;
101 int len;
102
103 /* logname consists of "fsname-nodetype".
104 * e.g. "lustre-MDT0001", "SUN-000-client"
105 * there is an exception: llog "params" */
106 name_end = strrchr(logname, '-');
107 if (!name_end)
108 len = strlen(logname);
109 else
110 len = name_end - logname;
111 return mgc_name2resid(logname, len, res_id, type);
112 }
113
114 /********************** config llog list **********************/
115 static LIST_HEAD(config_llog_list);
116 static DEFINE_SPINLOCK(config_list_lock);
117
118 /* Take a reference to a config log */
config_log_get(struct config_llog_data * cld)119 static int config_log_get(struct config_llog_data *cld)
120 {
121 atomic_inc(&cld->cld_refcount);
122 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
123 atomic_read(&cld->cld_refcount));
124 return 0;
125 }
126
127 /* Drop a reference to a config log. When no longer referenced,
128 we can free the config log data */
config_log_put(struct config_llog_data * cld)129 static void config_log_put(struct config_llog_data *cld)
130 {
131 CDEBUG(D_INFO, "log %s refs %d\n", cld->cld_logname,
132 atomic_read(&cld->cld_refcount));
133 LASSERT(atomic_read(&cld->cld_refcount) > 0);
134
135 /* spinlock to make sure no item with 0 refcount in the list */
136 if (atomic_dec_and_lock(&cld->cld_refcount, &config_list_lock)) {
137 list_del(&cld->cld_list_chain);
138 spin_unlock(&config_list_lock);
139
140 CDEBUG(D_MGC, "dropping config log %s\n", cld->cld_logname);
141
142 if (cld->cld_recover)
143 config_log_put(cld->cld_recover);
144 if (cld->cld_sptlrpc)
145 config_log_put(cld->cld_sptlrpc);
146 if (cld->cld_params)
147 config_log_put(cld->cld_params);
148 if (cld_is_sptlrpc(cld))
149 sptlrpc_conf_log_stop(cld->cld_logname);
150
151 class_export_put(cld->cld_mgcexp);
152 kfree(cld);
153 }
154 }
155
156 /* Find a config log by name */
157 static
config_log_find(char * logname,struct config_llog_instance * cfg)158 struct config_llog_data *config_log_find(char *logname,
159 struct config_llog_instance *cfg)
160 {
161 struct config_llog_data *cld;
162 struct config_llog_data *found = NULL;
163 void *instance;
164
165 LASSERT(logname != NULL);
166
167 instance = cfg ? cfg->cfg_instance : NULL;
168 spin_lock(&config_list_lock);
169 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
170 /* check if instance equals */
171 if (instance != cld->cld_cfg.cfg_instance)
172 continue;
173
174 /* instance may be NULL, should check name */
175 if (strcmp(logname, cld->cld_logname) == 0) {
176 found = cld;
177 break;
178 }
179 }
180 if (found) {
181 atomic_inc(&found->cld_refcount);
182 LASSERT(found->cld_stopping == 0 || cld_is_sptlrpc(found) == 0);
183 }
184 spin_unlock(&config_list_lock);
185 return found;
186 }
187
188 static
do_config_log_add(struct obd_device * obd,char * logname,int type,struct config_llog_instance * cfg,struct super_block * sb)189 struct config_llog_data *do_config_log_add(struct obd_device *obd,
190 char *logname,
191 int type,
192 struct config_llog_instance *cfg,
193 struct super_block *sb)
194 {
195 struct config_llog_data *cld;
196 int rc;
197
198 CDEBUG(D_MGC, "do adding config log %s:%p\n", logname,
199 cfg ? cfg->cfg_instance : NULL);
200
201 cld = kzalloc(sizeof(*cld) + strlen(logname) + 1, GFP_NOFS);
202 if (!cld)
203 return ERR_PTR(-ENOMEM);
204
205 strcpy(cld->cld_logname, logname);
206 if (cfg)
207 cld->cld_cfg = *cfg;
208 else
209 cld->cld_cfg.cfg_callback = class_config_llog_handler;
210 mutex_init(&cld->cld_lock);
211 cld->cld_cfg.cfg_last_idx = 0;
212 cld->cld_cfg.cfg_flags = 0;
213 cld->cld_cfg.cfg_sb = sb;
214 cld->cld_type = type;
215 atomic_set(&cld->cld_refcount, 1);
216
217 /* Keep the mgc around until we are done */
218 cld->cld_mgcexp = class_export_get(obd->obd_self_export);
219
220 if (cld_is_sptlrpc(cld)) {
221 sptlrpc_conf_log_start(logname);
222 cld->cld_cfg.cfg_obdname = obd->obd_name;
223 }
224
225 rc = mgc_logname2resid(logname, &cld->cld_resid, type);
226
227 spin_lock(&config_list_lock);
228 list_add(&cld->cld_list_chain, &config_llog_list);
229 spin_unlock(&config_list_lock);
230
231 if (rc) {
232 config_log_put(cld);
233 return ERR_PTR(rc);
234 }
235
236 if (cld_is_sptlrpc(cld)) {
237 rc = mgc_process_log(obd, cld);
238 if (rc && rc != -ENOENT)
239 CERROR("failed processing sptlrpc log: %d\n", rc);
240 }
241
242 return cld;
243 }
244
config_recover_log_add(struct obd_device * obd,char * fsname,struct config_llog_instance * cfg,struct super_block * sb)245 static struct config_llog_data *config_recover_log_add(struct obd_device *obd,
246 char *fsname,
247 struct config_llog_instance *cfg,
248 struct super_block *sb)
249 {
250 struct config_llog_instance lcfg = *cfg;
251 struct config_llog_data *cld;
252 char logname[32];
253
254 /* we have to use different llog for clients and mdts for cmd
255 * where only clients are notified if one of cmd server restarts */
256 LASSERT(strlen(fsname) < sizeof(logname) / 2);
257 strcpy(logname, fsname);
258 LASSERT(lcfg.cfg_instance);
259 strcat(logname, "-cliir");
260
261 cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
262 return cld;
263 }
264
config_params_log_add(struct obd_device * obd,struct config_llog_instance * cfg,struct super_block * sb)265 static struct config_llog_data *config_params_log_add(struct obd_device *obd,
266 struct config_llog_instance *cfg, struct super_block *sb)
267 {
268 struct config_llog_instance lcfg = *cfg;
269 struct config_llog_data *cld;
270
271 lcfg.cfg_instance = sb;
272
273 cld = do_config_log_add(obd, PARAMS_FILENAME, CONFIG_T_PARAMS,
274 &lcfg, sb);
275
276 return cld;
277 }
278
279 /** Add this log to the list of active logs watched by an MGC.
280 * Active means we're watching for updates.
281 * We have one active log per "mount" - client instance or servername.
282 * Each instance may be at a different point in the log.
283 */
config_log_add(struct obd_device * obd,char * logname,struct config_llog_instance * cfg,struct super_block * sb)284 static int config_log_add(struct obd_device *obd, char *logname,
285 struct config_llog_instance *cfg,
286 struct super_block *sb)
287 {
288 struct lustre_sb_info *lsi = s2lsi(sb);
289 struct config_llog_data *cld;
290 struct config_llog_data *sptlrpc_cld;
291 struct config_llog_data *params_cld;
292 char seclogname[32];
293 char *ptr;
294 int rc;
295
296 CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
297
298 /*
299 * for each regular log, the depended sptlrpc log name is
300 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
301 */
302 ptr = strrchr(logname, '-');
303 if (ptr == NULL || ptr - logname > 8) {
304 CERROR("logname %s is too long\n", logname);
305 return -EINVAL;
306 }
307
308 memcpy(seclogname, logname, ptr - logname);
309 strcpy(seclogname + (ptr - logname), "-sptlrpc");
310
311 sptlrpc_cld = config_log_find(seclogname, NULL);
312 if (sptlrpc_cld == NULL) {
313 sptlrpc_cld = do_config_log_add(obd, seclogname,
314 CONFIG_T_SPTLRPC, NULL, NULL);
315 if (IS_ERR(sptlrpc_cld)) {
316 CERROR("can't create sptlrpc log: %s\n", seclogname);
317 rc = PTR_ERR(sptlrpc_cld);
318 goto out_err;
319 }
320 }
321 params_cld = config_params_log_add(obd, cfg, sb);
322 if (IS_ERR(params_cld)) {
323 rc = PTR_ERR(params_cld);
324 CERROR("%s: can't create params log: rc = %d\n",
325 obd->obd_name, rc);
326 goto out_err1;
327 }
328
329 cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
330 if (IS_ERR(cld)) {
331 CERROR("can't create log: %s\n", logname);
332 rc = PTR_ERR(cld);
333 goto out_err2;
334 }
335
336 cld->cld_sptlrpc = sptlrpc_cld;
337 cld->cld_params = params_cld;
338
339 LASSERT(lsi->lsi_lmd);
340 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
341 struct config_llog_data *recover_cld;
342 *strrchr(seclogname, '-') = 0;
343 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
344 if (IS_ERR(recover_cld)) {
345 rc = PTR_ERR(recover_cld);
346 goto out_err3;
347 }
348 cld->cld_recover = recover_cld;
349 }
350
351 return 0;
352
353 out_err3:
354 config_log_put(cld);
355
356 out_err2:
357 config_log_put(params_cld);
358
359 out_err1:
360 config_log_put(sptlrpc_cld);
361
362 out_err:
363 return rc;
364 }
365
366 DEFINE_MUTEX(llog_process_lock);
367
368 /** Stop watching for updates on this log.
369 */
config_log_end(char * logname,struct config_llog_instance * cfg)370 static int config_log_end(char *logname, struct config_llog_instance *cfg)
371 {
372 struct config_llog_data *cld;
373 struct config_llog_data *cld_sptlrpc = NULL;
374 struct config_llog_data *cld_params = NULL;
375 struct config_llog_data *cld_recover = NULL;
376 int rc = 0;
377
378 cld = config_log_find(logname, cfg);
379 if (cld == NULL)
380 return -ENOENT;
381
382 mutex_lock(&cld->cld_lock);
383 /*
384 * if cld_stopping is set, it means we didn't start the log thus
385 * not owning the start ref. this can happen after previous umount:
386 * the cld still hanging there waiting for lock cancel, and we
387 * remount again but failed in the middle and call log_end without
388 * calling start_log.
389 */
390 if (unlikely(cld->cld_stopping)) {
391 mutex_unlock(&cld->cld_lock);
392 /* drop the ref from the find */
393 config_log_put(cld);
394 return rc;
395 }
396
397 cld->cld_stopping = 1;
398
399 cld_recover = cld->cld_recover;
400 cld->cld_recover = NULL;
401 mutex_unlock(&cld->cld_lock);
402
403 if (cld_recover) {
404 mutex_lock(&cld_recover->cld_lock);
405 cld_recover->cld_stopping = 1;
406 mutex_unlock(&cld_recover->cld_lock);
407 config_log_put(cld_recover);
408 }
409
410 spin_lock(&config_list_lock);
411 cld_sptlrpc = cld->cld_sptlrpc;
412 cld->cld_sptlrpc = NULL;
413 cld_params = cld->cld_params;
414 cld->cld_params = NULL;
415 spin_unlock(&config_list_lock);
416
417 if (cld_sptlrpc)
418 config_log_put(cld_sptlrpc);
419
420 if (cld_params) {
421 mutex_lock(&cld_params->cld_lock);
422 cld_params->cld_stopping = 1;
423 mutex_unlock(&cld_params->cld_lock);
424 config_log_put(cld_params);
425 }
426
427 /* drop the ref from the find */
428 config_log_put(cld);
429 /* drop the start ref */
430 config_log_put(cld);
431
432 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
433 rc);
434 return rc;
435 }
436
lprocfs_mgc_rd_ir_state(struct seq_file * m,void * data)437 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
438 {
439 struct obd_device *obd = data;
440 struct obd_import *imp;
441 struct obd_connect_data *ocd;
442 struct config_llog_data *cld;
443 int rc;
444
445 rc = lprocfs_climp_check(obd);
446 if (rc)
447 return rc;
448
449 imp = obd->u.cli.cl_import;
450 ocd = &imp->imp_connect_data;
451
452 seq_printf(m, "imperative_recovery: %s\n",
453 OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
454 seq_printf(m, "client_state:\n");
455
456 spin_lock(&config_list_lock);
457 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
458 if (cld->cld_recover == NULL)
459 continue;
460 seq_printf(m, " - { client: %s, nidtbl_version: %u }\n",
461 cld->cld_logname,
462 cld->cld_recover->cld_cfg.cfg_last_idx);
463 }
464 spin_unlock(&config_list_lock);
465
466 LPROCFS_CLIMP_EXIT(obd);
467 return 0;
468 }
469
470 /* reenqueue any lost locks */
471 #define RQ_RUNNING 0x1
472 #define RQ_NOW 0x2
473 #define RQ_LATER 0x4
474 #define RQ_STOP 0x8
475 #define RQ_PRECLEANUP 0x10
476 static int rq_state;
477 static wait_queue_head_t rq_waitq;
478 static DECLARE_COMPLETION(rq_exit);
479 static DECLARE_COMPLETION(rq_start);
480
do_requeue(struct config_llog_data * cld)481 static void do_requeue(struct config_llog_data *cld)
482 {
483 LASSERT(atomic_read(&cld->cld_refcount) > 0);
484
485 /* Do not run mgc_process_log on a disconnected export or an
486 export which is being disconnected. Take the client
487 semaphore to make the check non-racy. */
488 down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
489 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
490 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
491 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
492 } else {
493 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
494 cld->cld_logname);
495 }
496 up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
497 }
498
499 /* this timeout represents how many seconds MGC should wait before
500 * requeue config and recover lock to the MGS. We need to randomize this
501 * in order to not flood the MGS.
502 */
503 #define MGC_TIMEOUT_MIN_SECONDS 5
504 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
505
mgc_requeue_thread(void * data)506 static int mgc_requeue_thread(void *data)
507 {
508 bool first = true;
509
510 CDEBUG(D_MGC, "Starting requeue thread\n");
511
512 /* Keep trying failed locks periodically */
513 spin_lock(&config_list_lock);
514 rq_state |= RQ_RUNNING;
515 while (1) {
516 struct l_wait_info lwi;
517 struct config_llog_data *cld, *cld_prev;
518 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
519 int stopped = !!(rq_state & RQ_STOP);
520 int to;
521
522 /* Any new or requeued lostlocks will change the state */
523 rq_state &= ~(RQ_NOW | RQ_LATER);
524 spin_unlock(&config_list_lock);
525
526 if (first) {
527 first = false;
528 complete(&rq_start);
529 }
530
531 /* Always wait a few seconds to allow the server who
532 caused the lock revocation to finish its setup, plus some
533 random so everyone doesn't try to reconnect at once. */
534 to = MGC_TIMEOUT_MIN_SECONDS * HZ;
535 to += rand * HZ / 100; /* rand is centi-seconds */
536 lwi = LWI_TIMEOUT(to, NULL, NULL);
537 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
538 &lwi);
539
540 /*
541 * iterate & processing through the list. for each cld, process
542 * its depending sptlrpc cld firstly (if any) and then itself.
543 *
544 * it's guaranteed any item in the list must have
545 * reference > 0; and if cld_lostlock is set, at
546 * least one reference is taken by the previous enqueue.
547 */
548 cld_prev = NULL;
549
550 spin_lock(&config_list_lock);
551 rq_state &= ~RQ_PRECLEANUP;
552 list_for_each_entry(cld, &config_llog_list,
553 cld_list_chain) {
554 if (!cld->cld_lostlock)
555 continue;
556
557 spin_unlock(&config_list_lock);
558
559 LASSERT(atomic_read(&cld->cld_refcount) > 0);
560
561 /* Whether we enqueued again or not in mgc_process_log,
562 * we're done with the ref from the old enqueue */
563 if (cld_prev)
564 config_log_put(cld_prev);
565 cld_prev = cld;
566
567 cld->cld_lostlock = 0;
568 if (likely(!stopped))
569 do_requeue(cld);
570
571 spin_lock(&config_list_lock);
572 }
573 spin_unlock(&config_list_lock);
574 if (cld_prev)
575 config_log_put(cld_prev);
576
577 /* break after scanning the list so that we can drop
578 * refcount to losing lock clds */
579 if (unlikely(stopped)) {
580 spin_lock(&config_list_lock);
581 break;
582 }
583
584 /* Wait a bit to see if anyone else needs a requeue */
585 lwi = (struct l_wait_info) { 0 };
586 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
587 &lwi);
588 spin_lock(&config_list_lock);
589 }
590 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
591 rq_state &= ~RQ_RUNNING;
592 spin_unlock(&config_list_lock);
593
594 complete(&rq_exit);
595
596 CDEBUG(D_MGC, "Ending requeue thread\n");
597 return 0;
598 }
599
600 /* Add a cld to the list to requeue. Start the requeue thread if needed.
601 We are responsible for dropping the config log reference from here on out. */
mgc_requeue_add(struct config_llog_data * cld)602 static void mgc_requeue_add(struct config_llog_data *cld)
603 {
604 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
605 cld->cld_logname, atomic_read(&cld->cld_refcount),
606 cld->cld_stopping, rq_state);
607 LASSERT(atomic_read(&cld->cld_refcount) > 0);
608
609 mutex_lock(&cld->cld_lock);
610 if (cld->cld_stopping || cld->cld_lostlock) {
611 mutex_unlock(&cld->cld_lock);
612 return;
613 }
614 /* this refcount will be released in mgc_requeue_thread. */
615 config_log_get(cld);
616 cld->cld_lostlock = 1;
617 mutex_unlock(&cld->cld_lock);
618
619 /* Hold lock for rq_state */
620 spin_lock(&config_list_lock);
621 if (rq_state & RQ_STOP) {
622 spin_unlock(&config_list_lock);
623 cld->cld_lostlock = 0;
624 config_log_put(cld);
625 } else {
626 rq_state |= RQ_NOW;
627 spin_unlock(&config_list_lock);
628 wake_up(&rq_waitq);
629 }
630 }
631
mgc_llog_init(const struct lu_env * env,struct obd_device * obd)632 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
633 {
634 struct llog_ctxt *ctxt;
635 int rc;
636
637 /* setup only remote ctxt, the local disk context is switched per each
638 * filesystem during mgc_fs_setup() */
639 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
640 &llog_client_ops);
641 if (rc)
642 return rc;
643
644 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
645 LASSERT(ctxt);
646
647 llog_initiator_connect(ctxt);
648 llog_ctxt_put(ctxt);
649
650 return 0;
651 }
652
mgc_llog_fini(const struct lu_env * env,struct obd_device * obd)653 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
654 {
655 struct llog_ctxt *ctxt;
656
657 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
658 if (ctxt)
659 llog_cleanup(env, ctxt);
660
661 return 0;
662 }
663
664 static atomic_t mgc_count = ATOMIC_INIT(0);
mgc_precleanup(struct obd_device * obd,enum obd_cleanup_stage stage)665 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
666 {
667 int rc = 0;
668 int temp;
669
670 switch (stage) {
671 case OBD_CLEANUP_EARLY:
672 break;
673 case OBD_CLEANUP_EXPORTS:
674 if (atomic_dec_and_test(&mgc_count)) {
675 LASSERT(rq_state & RQ_RUNNING);
676 /* stop requeue thread */
677 temp = RQ_STOP;
678 } else {
679 /* wakeup requeue thread to clean our cld */
680 temp = RQ_NOW | RQ_PRECLEANUP;
681 }
682 spin_lock(&config_list_lock);
683 rq_state |= temp;
684 spin_unlock(&config_list_lock);
685 wake_up(&rq_waitq);
686 if (temp & RQ_STOP)
687 wait_for_completion(&rq_exit);
688 obd_cleanup_client_import(obd);
689 rc = mgc_llog_fini(NULL, obd);
690 if (rc != 0)
691 CERROR("failed to cleanup llogging subsystems\n");
692 break;
693 }
694 return rc;
695 }
696
mgc_cleanup(struct obd_device * obd)697 static int mgc_cleanup(struct obd_device *obd)
698 {
699 /* COMPAT_146 - old config logs may have added profiles we don't
700 know about */
701 if (obd->obd_type->typ_refcnt <= 1)
702 /* Only for the last mgc */
703 class_del_profiles();
704
705 lprocfs_obd_cleanup(obd);
706 ptlrpcd_decref();
707
708 return client_obd_cleanup(obd);
709 }
710
mgc_setup(struct obd_device * obd,struct lustre_cfg * lcfg)711 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
712 {
713 struct lprocfs_static_vars lvars = { NULL };
714 int rc;
715
716 ptlrpcd_addref();
717
718 rc = client_obd_setup(obd, lcfg);
719 if (rc)
720 goto err_decref;
721
722 rc = mgc_llog_init(NULL, obd);
723 if (rc) {
724 CERROR("failed to setup llogging subsystems\n");
725 goto err_cleanup;
726 }
727
728 lprocfs_mgc_init_vars(&lvars);
729 lprocfs_obd_setup(obd, lvars.obd_vars, lvars.sysfs_vars);
730 sptlrpc_lprocfs_cliobd_attach(obd);
731
732 if (atomic_inc_return(&mgc_count) == 1) {
733 rq_state = 0;
734 init_waitqueue_head(&rq_waitq);
735
736 /* start requeue thread */
737 rc = PTR_ERR(kthread_run(mgc_requeue_thread, NULL,
738 "ll_cfg_requeue"));
739 if (IS_ERR_VALUE(rc)) {
740 CERROR("%s: Cannot start requeue thread (%d),no more log updates!\n",
741 obd->obd_name, rc);
742 goto err_cleanup;
743 }
744 /* rc is the task_struct pointer of mgc_requeue_thread. */
745 rc = 0;
746 wait_for_completion(&rq_start);
747 }
748
749 return rc;
750
751 err_cleanup:
752 client_obd_cleanup(obd);
753 err_decref:
754 ptlrpcd_decref();
755 return rc;
756 }
757
758 /* based on ll_mdc_blocking_ast */
mgc_blocking_ast(struct ldlm_lock * lock,struct ldlm_lock_desc * desc,void * data,int flag)759 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
760 void *data, int flag)
761 {
762 struct lustre_handle lockh;
763 struct config_llog_data *cld = data;
764 int rc = 0;
765
766 switch (flag) {
767 case LDLM_CB_BLOCKING:
768 /* mgs wants the lock, give it up... */
769 LDLM_DEBUG(lock, "MGC blocking CB");
770 ldlm_lock2handle(lock, &lockh);
771 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
772 break;
773 case LDLM_CB_CANCELING:
774 /* We've given up the lock, prepare ourselves to update. */
775 LDLM_DEBUG(lock, "MGC cancel CB");
776
777 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
778 PLDLMRES(lock->l_resource),
779 (char *)&lock->l_resource->lr_name.name[0]);
780
781 if (!cld) {
782 CDEBUG(D_INFO, "missing data, won't requeue\n");
783 break;
784 }
785
786 /* held at mgc_process_log(). */
787 LASSERT(atomic_read(&cld->cld_refcount) > 0);
788 /* Are we done with this log? */
789 if (cld->cld_stopping) {
790 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
791 cld->cld_logname);
792 config_log_put(cld);
793 break;
794 }
795 /* Make sure not to re-enqueue when the mgc is stopping
796 (we get called from client_disconnect_export) */
797 if (!lock->l_conn_export ||
798 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
799 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
800 cld->cld_logname);
801 config_log_put(cld);
802 break;
803 }
804
805 /* Re-enqueue now */
806 mgc_requeue_add(cld);
807 config_log_put(cld);
808 break;
809 default:
810 LBUG();
811 }
812
813 return rc;
814 }
815
816 /* Not sure where this should go... */
817 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
818 * that we can have a chance to try the secondary MGS if any. */
819 #define MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
820 + PING_INTERVAL)
821 #define MGC_TARGET_REG_LIMIT 10
822 #define MGC_SEND_PARAM_LIMIT 10
823
824 /* Send parameter to MGS*/
mgc_set_mgs_param(struct obd_export * exp,struct mgs_send_param * msp)825 static int mgc_set_mgs_param(struct obd_export *exp,
826 struct mgs_send_param *msp)
827 {
828 struct ptlrpc_request *req;
829 struct mgs_send_param *req_msp, *rep_msp;
830 int rc;
831
832 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
833 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
834 MGS_SET_INFO);
835 if (!req)
836 return -ENOMEM;
837
838 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
839 if (!req_msp) {
840 ptlrpc_req_finished(req);
841 return -ENOMEM;
842 }
843
844 memcpy(req_msp, msp, sizeof(*req_msp));
845 ptlrpc_request_set_replen(req);
846
847 /* Limit how long we will wait for the enqueue to complete */
848 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
849 rc = ptlrpc_queue_wait(req);
850 if (!rc) {
851 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
852 memcpy(msp, rep_msp, sizeof(*rep_msp));
853 }
854
855 ptlrpc_req_finished(req);
856
857 return rc;
858 }
859
860 /* Take a config lock so we can get cancel notifications */
mgc_enqueue(struct obd_export * exp,struct lov_stripe_md * lsm,__u32 type,ldlm_policy_data_t * policy,__u32 mode,__u64 * flags,void * bl_cb,void * cp_cb,void * gl_cb,void * data,__u32 lvb_len,void * lvb_swabber,struct lustre_handle * lockh)861 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
862 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
863 __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
864 void *data, __u32 lvb_len, void *lvb_swabber,
865 struct lustre_handle *lockh)
866 {
867 struct config_llog_data *cld = data;
868 struct ldlm_enqueue_info einfo = {
869 .ei_type = type,
870 .ei_mode = mode,
871 .ei_cb_bl = mgc_blocking_ast,
872 .ei_cb_cp = ldlm_completion_ast,
873 };
874 struct ptlrpc_request *req;
875 int short_limit = cld_is_sptlrpc(cld);
876 int rc;
877
878 CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
879 cld->cld_resid.name[0]);
880
881 /* We need a callback for every lockholder, so don't try to
882 ldlm_lock_match (see rev 1.1.2.11.2.47) */
883 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
884 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
885 LDLM_ENQUEUE);
886 if (req == NULL)
887 return -ENOMEM;
888
889 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
890 ptlrpc_request_set_replen(req);
891
892 /* Limit how long we will wait for the enqueue to complete */
893 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
894 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
895 NULL, 0, LVB_T_NONE, lockh, 0);
896 /* A failed enqueue should still call the mgc_blocking_ast,
897 where it will be requeued if needed ("grant failed"). */
898 ptlrpc_req_finished(req);
899 return rc;
900 }
901
mgc_notify_active(struct obd_device * unused)902 static void mgc_notify_active(struct obd_device *unused)
903 {
904 /* wakeup mgc_requeue_thread to requeue mgc lock */
905 spin_lock(&config_list_lock);
906 rq_state |= RQ_NOW;
907 spin_unlock(&config_list_lock);
908 wake_up(&rq_waitq);
909
910 /* TODO: Help the MGS rebuild nidtbl. -jay */
911 }
912
913 /* Send target_reg message to MGS */
mgc_target_register(struct obd_export * exp,struct mgs_target_info * mti)914 static int mgc_target_register(struct obd_export *exp,
915 struct mgs_target_info *mti)
916 {
917 struct ptlrpc_request *req;
918 struct mgs_target_info *req_mti, *rep_mti;
919 int rc;
920
921 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
922 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
923 MGS_TARGET_REG);
924 if (req == NULL)
925 return -ENOMEM;
926
927 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
928 if (!req_mti) {
929 ptlrpc_req_finished(req);
930 return -ENOMEM;
931 }
932
933 memcpy(req_mti, mti, sizeof(*req_mti));
934 ptlrpc_request_set_replen(req);
935 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
936 /* Limit how long we will wait for the enqueue to complete */
937 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
938
939 rc = ptlrpc_queue_wait(req);
940 if (!rc) {
941 rep_mti = req_capsule_server_get(&req->rq_pill,
942 &RMF_MGS_TARGET_INFO);
943 memcpy(mti, rep_mti, sizeof(*rep_mti));
944 CDEBUG(D_MGC, "register %s got index = %d\n",
945 mti->mti_svname, mti->mti_stripe_index);
946 }
947 ptlrpc_req_finished(req);
948
949 return rc;
950 }
951
mgc_set_info_async(const struct lu_env * env,struct obd_export * exp,u32 keylen,void * key,u32 vallen,void * val,struct ptlrpc_request_set * set)952 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
953 u32 keylen, void *key, u32 vallen,
954 void *val, struct ptlrpc_request_set *set)
955 {
956 int rc = -EINVAL;
957
958 /* Turn off initial_recov after we try all backup servers once */
959 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
960 struct obd_import *imp = class_exp2cliimp(exp);
961 int value;
962
963 if (vallen != sizeof(int))
964 return -EINVAL;
965 value = *(int *)val;
966 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
967 imp->imp_obd->obd_name, value,
968 imp->imp_deactive, imp->imp_invalid,
969 imp->imp_replayable, imp->imp_obd->obd_replayable,
970 ptlrpc_import_state_name(imp->imp_state));
971 /* Resurrect if we previously died */
972 if ((imp->imp_state != LUSTRE_IMP_FULL &&
973 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
974 ptlrpc_reconnect_import(imp);
975 return 0;
976 }
977 if (KEY_IS(KEY_SET_INFO)) {
978 struct mgs_send_param *msp;
979
980 msp = val;
981 rc = mgc_set_mgs_param(exp, msp);
982 return rc;
983 }
984 if (KEY_IS(KEY_MGSSEC)) {
985 struct client_obd *cli = &exp->exp_obd->u.cli;
986 struct sptlrpc_flavor flvr;
987
988 /*
989 * empty string means using current flavor, if which haven't
990 * been set yet, set it as null.
991 *
992 * if flavor has been set previously, check the asking flavor
993 * must match the existing one.
994 */
995 if (vallen == 0) {
996 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
997 return 0;
998 val = "null";
999 vallen = 4;
1000 }
1001
1002 rc = sptlrpc_parse_flavor(val, &flvr);
1003 if (rc) {
1004 CERROR("invalid sptlrpc flavor %s to MGS\n",
1005 (char *) val);
1006 return rc;
1007 }
1008
1009 /*
1010 * caller already hold a mutex
1011 */
1012 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1013 cli->cl_flvr_mgc = flvr;
1014 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1015 sizeof(flvr)) != 0) {
1016 char str[20];
1017
1018 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1019 str, sizeof(str));
1020 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but currently %s is in use\n",
1021 (char *) val, str);
1022 rc = -EPERM;
1023 }
1024 return rc;
1025 }
1026
1027 return rc;
1028 }
1029
mgc_get_info(const struct lu_env * env,struct obd_export * exp,__u32 keylen,void * key,__u32 * vallen,void * val,struct lov_stripe_md * unused)1030 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1031 __u32 keylen, void *key, __u32 *vallen, void *val,
1032 struct lov_stripe_md *unused)
1033 {
1034 int rc = -EINVAL;
1035
1036 if (KEY_IS(KEY_CONN_DATA)) {
1037 struct obd_import *imp = class_exp2cliimp(exp);
1038 struct obd_connect_data *data = val;
1039
1040 if (*vallen == sizeof(*data)) {
1041 *data = imp->imp_connect_data;
1042 rc = 0;
1043 }
1044 }
1045
1046 return rc;
1047 }
1048
mgc_import_event(struct obd_device * obd,struct obd_import * imp,enum obd_import_event event)1049 static int mgc_import_event(struct obd_device *obd,
1050 struct obd_import *imp,
1051 enum obd_import_event event)
1052 {
1053 LASSERT(imp->imp_obd == obd);
1054 CDEBUG(D_MGC, "import event %#x\n", event);
1055
1056 switch (event) {
1057 case IMP_EVENT_DISCON:
1058 /* MGC imports should not wait for recovery */
1059 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1060 ptlrpc_pinger_ir_down();
1061 break;
1062 case IMP_EVENT_INACTIVE:
1063 break;
1064 case IMP_EVENT_INVALIDATE: {
1065 struct ldlm_namespace *ns = obd->obd_namespace;
1066
1067 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1068 break;
1069 }
1070 case IMP_EVENT_ACTIVE:
1071 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1072 /* Clearing obd_no_recov allows us to continue pinging */
1073 obd->obd_no_recov = 0;
1074 mgc_notify_active(obd);
1075 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1076 ptlrpc_pinger_ir_up();
1077 break;
1078 case IMP_EVENT_OCD:
1079 break;
1080 case IMP_EVENT_DEACTIVATE:
1081 case IMP_EVENT_ACTIVATE:
1082 break;
1083 default:
1084 CERROR("Unknown import event %#x\n", event);
1085 LBUG();
1086 }
1087 return 0;
1088 }
1089
1090 enum {
1091 CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_CACHE_SHIFT),
1092 CONFIG_READ_NRPAGES = 4
1093 };
1094
mgc_apply_recover_logs(struct obd_device * mgc,struct config_llog_data * cld,__u64 max_version,void * data,int datalen,bool mne_swab)1095 static int mgc_apply_recover_logs(struct obd_device *mgc,
1096 struct config_llog_data *cld,
1097 __u64 max_version,
1098 void *data, int datalen, bool mne_swab)
1099 {
1100 struct config_llog_instance *cfg = &cld->cld_cfg;
1101 struct mgs_nidtbl_entry *entry;
1102 struct lustre_cfg *lcfg;
1103 struct lustre_cfg_bufs bufs;
1104 u64 prev_version = 0;
1105 char *inst;
1106 char *buf;
1107 int bufsz;
1108 int pos;
1109 int rc = 0;
1110 int off = 0;
1111
1112 LASSERT(cfg->cfg_instance != NULL);
1113 LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1114
1115 inst = kzalloc(PAGE_CACHE_SIZE, GFP_KERNEL);
1116 if (!inst)
1117 return -ENOMEM;
1118
1119 pos = snprintf(inst, PAGE_CACHE_SIZE, "%p", cfg->cfg_instance);
1120 if (pos >= PAGE_CACHE_SIZE) {
1121 kfree(inst);
1122 return -E2BIG;
1123 }
1124
1125 ++pos;
1126 buf = inst + pos;
1127 bufsz = PAGE_CACHE_SIZE - pos;
1128
1129 while (datalen > 0) {
1130 int entry_len = sizeof(*entry);
1131 int is_ost;
1132 struct obd_device *obd;
1133 char *obdname;
1134 char *cname;
1135 char *params;
1136 char *uuid;
1137
1138 rc = -EINVAL;
1139 if (datalen < sizeof(*entry))
1140 break;
1141
1142 entry = (typeof(entry))(data + off);
1143
1144 /* sanity check */
1145 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1146 break;
1147 if (entry->mne_nid_count == 0) /* at least one nid entry */
1148 break;
1149 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1150 break;
1151
1152 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1153 if (datalen < entry_len) /* must have entry_len at least */
1154 break;
1155
1156 /* Keep this swab for normal mixed endian handling. LU-1644 */
1157 if (mne_swab)
1158 lustre_swab_mgs_nidtbl_entry(entry);
1159 if (entry->mne_length > PAGE_CACHE_SIZE) {
1160 CERROR("MNE too large (%u)\n", entry->mne_length);
1161 break;
1162 }
1163
1164 if (entry->mne_length < entry_len)
1165 break;
1166
1167 off += entry->mne_length;
1168 datalen -= entry->mne_length;
1169 if (datalen < 0)
1170 break;
1171
1172 if (entry->mne_version > max_version) {
1173 CERROR("entry index(%lld) is over max_index(%lld)\n",
1174 entry->mne_version, max_version);
1175 break;
1176 }
1177
1178 if (prev_version >= entry->mne_version) {
1179 CERROR("index unsorted, prev %lld, now %lld\n",
1180 prev_version, entry->mne_version);
1181 break;
1182 }
1183 prev_version = entry->mne_version;
1184
1185 /*
1186 * Write a string with format "nid::instance" to
1187 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1188 */
1189
1190 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1191 memset(buf, 0, bufsz);
1192 obdname = buf;
1193 pos = 0;
1194
1195 /* lustre-OST0001-osc-<instance #> */
1196 strcpy(obdname, cld->cld_logname);
1197 cname = strrchr(obdname, '-');
1198 if (cname == NULL) {
1199 CERROR("mgc %s: invalid logname %s\n",
1200 mgc->obd_name, obdname);
1201 break;
1202 }
1203
1204 pos = cname - obdname;
1205 obdname[pos] = 0;
1206 pos += sprintf(obdname + pos, "-%s%04x",
1207 is_ost ? "OST" : "MDT", entry->mne_index);
1208
1209 cname = is_ost ? "osc" : "mdc";
1210 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1211 lustre_cfg_bufs_reset(&bufs, obdname);
1212
1213 /* find the obd by obdname */
1214 obd = class_name2obd(obdname);
1215 if (obd == NULL) {
1216 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1217 mgc->obd_name, obdname);
1218 rc = 0;
1219 /* this is a safe race, when the ost is starting up...*/
1220 continue;
1221 }
1222
1223 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1224 ++pos;
1225 params = buf + pos;
1226 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1227 uuid = buf + pos;
1228
1229 down_read(&obd->u.cli.cl_sem);
1230 if (obd->u.cli.cl_import == NULL) {
1231 /* client does not connect to the OST yet */
1232 up_read(&obd->u.cli.cl_sem);
1233 rc = 0;
1234 continue;
1235 }
1236
1237 /* TODO: iterate all nids to find one */
1238 /* find uuid by nid */
1239 rc = client_import_find_conn(obd->u.cli.cl_import,
1240 entry->u.nids[0],
1241 (struct obd_uuid *)uuid);
1242 up_read(&obd->u.cli.cl_sem);
1243 if (rc < 0) {
1244 CERROR("mgc: cannot find uuid by nid %s\n",
1245 libcfs_nid2str(entry->u.nids[0]));
1246 break;
1247 }
1248
1249 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1250 uuid, libcfs_nid2str(entry->u.nids[0]));
1251
1252 pos += strlen(uuid);
1253 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1254 LASSERT(pos < bufsz);
1255
1256 lustre_cfg_bufs_set_string(&bufs, 1, params);
1257
1258 rc = -ENOMEM;
1259 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1260 if (lcfg == NULL) {
1261 CERROR("mgc: cannot allocate memory\n");
1262 break;
1263 }
1264
1265 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1266 prev_version, max_version, obdname, params);
1267
1268 rc = class_process_config(lcfg);
1269 lustre_cfg_free(lcfg);
1270 if (rc)
1271 CDEBUG(D_INFO, "process config for %s error %d\n",
1272 obdname, rc);
1273
1274 /* continue, even one with error */
1275 }
1276
1277 kfree(inst);
1278 return rc;
1279 }
1280
1281 /**
1282 * This function is called if this client was notified for target restarting
1283 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1284 */
mgc_process_recover_log(struct obd_device * obd,struct config_llog_data * cld)1285 static int mgc_process_recover_log(struct obd_device *obd,
1286 struct config_llog_data *cld)
1287 {
1288 struct ptlrpc_request *req = NULL;
1289 struct config_llog_instance *cfg = &cld->cld_cfg;
1290 struct mgs_config_body *body;
1291 struct mgs_config_res *res;
1292 struct ptlrpc_bulk_desc *desc;
1293 struct page **pages;
1294 int nrpages;
1295 bool eof = true;
1296 bool mne_swab = false;
1297 int i;
1298 int ealen;
1299 int rc;
1300
1301 /* allocate buffer for bulk transfer.
1302 * if this is the first time for this mgs to read logs,
1303 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1304 * once; otherwise, it only reads increment of logs, this should be
1305 * small and CONFIG_READ_NRPAGES will be used.
1306 */
1307 nrpages = CONFIG_READ_NRPAGES;
1308 if (cfg->cfg_last_idx == 0) /* the first time */
1309 nrpages = CONFIG_READ_NRPAGES_INIT;
1310
1311 pages = kcalloc(nrpages, sizeof(*pages), GFP_KERNEL);
1312 if (pages == NULL) {
1313 rc = -ENOMEM;
1314 goto out;
1315 }
1316
1317 for (i = 0; i < nrpages; i++) {
1318 pages[i] = alloc_page(GFP_KERNEL);
1319 if (pages[i] == NULL) {
1320 rc = -ENOMEM;
1321 goto out;
1322 }
1323 }
1324
1325 again:
1326 LASSERT(cld_is_recover(cld));
1327 LASSERT(mutex_is_locked(&cld->cld_lock));
1328 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1329 &RQF_MGS_CONFIG_READ);
1330 if (req == NULL) {
1331 rc = -ENOMEM;
1332 goto out;
1333 }
1334
1335 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1336 if (rc)
1337 goto out;
1338
1339 /* pack request */
1340 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1341 LASSERT(body != NULL);
1342 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1343 if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1344 >= sizeof(body->mcb_name)) {
1345 rc = -E2BIG;
1346 goto out;
1347 }
1348 body->mcb_offset = cfg->cfg_last_idx + 1;
1349 body->mcb_type = cld->cld_type;
1350 body->mcb_bits = PAGE_CACHE_SHIFT;
1351 body->mcb_units = nrpages;
1352
1353 /* allocate bulk transfer descriptor */
1354 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1, BULK_PUT_SINK,
1355 MGS_BULK_PORTAL);
1356 if (desc == NULL) {
1357 rc = -ENOMEM;
1358 goto out;
1359 }
1360
1361 for (i = 0; i < nrpages; i++)
1362 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
1363
1364 ptlrpc_request_set_replen(req);
1365 rc = ptlrpc_queue_wait(req);
1366 if (rc)
1367 goto out;
1368
1369 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1370 if (res->mcr_size < res->mcr_offset) {
1371 rc = -EINVAL;
1372 goto out;
1373 }
1374
1375 /* always update the index even though it might have errors with
1376 * handling the recover logs */
1377 cfg->cfg_last_idx = res->mcr_offset;
1378 eof = res->mcr_offset == res->mcr_size;
1379
1380 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1381 res->mcr_offset, eof == false);
1382
1383 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1384 if (ealen < 0) {
1385 rc = ealen;
1386 goto out;
1387 }
1388
1389 if (ealen > nrpages << PAGE_CACHE_SHIFT) {
1390 rc = -EINVAL;
1391 goto out;
1392 }
1393
1394 if (ealen == 0) { /* no logs transferred */
1395 if (!eof)
1396 rc = -EINVAL;
1397 goto out;
1398 }
1399
1400 mne_swab = !!ptlrpc_rep_need_swab(req);
1401 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1402 /* This import flag means the server did an extra swab of IR MNE
1403 * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1404 if (unlikely(req->rq_import->imp_need_mne_swab))
1405 mne_swab = !mne_swab;
1406 #else
1407 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1408 #endif
1409
1410 for (i = 0; i < nrpages && ealen > 0; i++) {
1411 int rc2;
1412 void *ptr;
1413
1414 ptr = kmap(pages[i]);
1415 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1416 min_t(int, ealen, PAGE_CACHE_SIZE),
1417 mne_swab);
1418 kunmap(pages[i]);
1419 if (rc2 < 0) {
1420 CWARN("Process recover log %s error %d\n",
1421 cld->cld_logname, rc2);
1422 break;
1423 }
1424
1425 ealen -= PAGE_CACHE_SIZE;
1426 }
1427
1428 out:
1429 if (req)
1430 ptlrpc_req_finished(req);
1431
1432 if (rc == 0 && !eof)
1433 goto again;
1434
1435 if (pages) {
1436 for (i = 0; i < nrpages; i++) {
1437 if (pages[i] == NULL)
1438 break;
1439 __free_page(pages[i]);
1440 }
1441 kfree(pages);
1442 }
1443 return rc;
1444 }
1445
1446 /* local_only means it cannot get remote llogs */
mgc_process_cfg_log(struct obd_device * mgc,struct config_llog_data * cld,int local_only)1447 static int mgc_process_cfg_log(struct obd_device *mgc,
1448 struct config_llog_data *cld, int local_only)
1449 {
1450 struct llog_ctxt *ctxt;
1451 struct lustre_sb_info *lsi = NULL;
1452 int rc = 0;
1453 bool sptlrpc_started = false;
1454 struct lu_env *env;
1455
1456 LASSERT(cld);
1457 LASSERT(mutex_is_locked(&cld->cld_lock));
1458
1459 /*
1460 * local copy of sptlrpc log is controlled elsewhere, don't try to
1461 * read it up here.
1462 */
1463 if (cld_is_sptlrpc(cld) && local_only)
1464 return 0;
1465
1466 if (cld->cld_cfg.cfg_sb)
1467 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1468
1469 env = kzalloc(sizeof(*env), GFP_KERNEL);
1470 if (!env)
1471 return -ENOMEM;
1472
1473 rc = lu_env_init(env, LCT_MG_THREAD);
1474 if (rc)
1475 goto out_free;
1476
1477 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1478 LASSERT(ctxt);
1479
1480 if (local_only) /* no local log at client side */ {
1481 rc = -EIO;
1482 goto out_pop;
1483 }
1484
1485 if (cld_is_sptlrpc(cld)) {
1486 sptlrpc_conf_log_update_begin(cld->cld_logname);
1487 sptlrpc_started = true;
1488 }
1489
1490 /* logname and instance info should be the same, so use our
1491 * copy of the instance for the update. The cfg_last_idx will
1492 * be updated here. */
1493 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1494 &cld->cld_cfg);
1495
1496 out_pop:
1497 __llog_ctxt_put(env, ctxt);
1498
1499 /*
1500 * update settings on existing OBDs. doing it inside
1501 * of llog_process_lock so no device is attaching/detaching
1502 * in parallel.
1503 * the logname must be <fsname>-sptlrpc
1504 */
1505 if (sptlrpc_started) {
1506 LASSERT(cld_is_sptlrpc(cld));
1507 sptlrpc_conf_log_update_end(cld->cld_logname);
1508 class_notify_sptlrpc_conf(cld->cld_logname,
1509 strlen(cld->cld_logname) -
1510 strlen("-sptlrpc"));
1511 }
1512
1513 lu_env_fini(env);
1514 out_free:
1515 kfree(env);
1516 return rc;
1517 }
1518
1519 /** Get a config log from the MGS and process it.
1520 * This func is called for both clients and servers.
1521 * Copy the log locally before parsing it if appropriate (non-MGS server)
1522 */
mgc_process_log(struct obd_device * mgc,struct config_llog_data * cld)1523 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1524 {
1525 struct lustre_handle lockh = { 0 };
1526 __u64 flags = LDLM_FL_NO_LRU;
1527 int rc = 0, rcl;
1528
1529 LASSERT(cld);
1530
1531 /* I don't want multiple processes running process_log at once --
1532 sounds like badness. It actually might be fine, as long as
1533 we're not trying to update from the same log
1534 simultaneously (in which case we should use a per-log sem.) */
1535 mutex_lock(&cld->cld_lock);
1536 if (cld->cld_stopping) {
1537 mutex_unlock(&cld->cld_lock);
1538 return 0;
1539 }
1540
1541 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1542
1543 CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1544 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1545
1546 /* Get the cfg lock on the llog */
1547 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1548 LCK_CR, &flags, NULL, NULL, NULL,
1549 cld, 0, NULL, &lockh);
1550 if (rcl == 0) {
1551 /* Get the cld, it will be released in mgc_blocking_ast. */
1552 config_log_get(cld);
1553 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1554 LASSERT(rc == 0);
1555 } else {
1556 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1557
1558 /* mark cld_lostlock so that it will requeue
1559 * after MGC becomes available. */
1560 cld->cld_lostlock = 1;
1561 /* Get extra reference, it will be put in requeue thread */
1562 config_log_get(cld);
1563 }
1564
1565 if (cld_is_recover(cld)) {
1566 rc = 0; /* this is not a fatal error for recover log */
1567 if (rcl == 0)
1568 rc = mgc_process_recover_log(mgc, cld);
1569 } else {
1570 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1571 }
1572
1573 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1574 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1575
1576 mutex_unlock(&cld->cld_lock);
1577
1578 /* Now drop the lock so MGS can revoke it */
1579 if (!rcl)
1580 ldlm_lock_decref(&lockh, LCK_CR);
1581
1582 return rc;
1583 }
1584
1585 /** Called from lustre_process_log.
1586 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1587 * any services, and adds it to the list logs to watch (follow).
1588 */
mgc_process_config(struct obd_device * obd,u32 len,void * buf)1589 static int mgc_process_config(struct obd_device *obd, u32 len, void *buf)
1590 {
1591 struct lustre_cfg *lcfg = buf;
1592 struct config_llog_instance *cfg = NULL;
1593 char *logname;
1594 int rc = 0;
1595
1596 switch (lcfg->lcfg_command) {
1597 case LCFG_LOV_ADD_OBD: {
1598 /* Overloading this cfg command: register a new target */
1599 struct mgs_target_info *mti;
1600
1601 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1602 sizeof(struct mgs_target_info)) {
1603 rc = -EINVAL;
1604 goto out;
1605 }
1606
1607 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1608 CDEBUG(D_MGC, "add_target %s %#x\n",
1609 mti->mti_svname, mti->mti_flags);
1610 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1611 break;
1612 }
1613 case LCFG_LOV_DEL_OBD:
1614 /* Unregister has no meaning at the moment. */
1615 CERROR("lov_del_obd unimplemented\n");
1616 rc = -ENOSYS;
1617 break;
1618 case LCFG_SPTLRPC_CONF: {
1619 rc = sptlrpc_process_config(lcfg);
1620 break;
1621 }
1622 case LCFG_LOG_START: {
1623 struct config_llog_data *cld;
1624 struct super_block *sb;
1625
1626 logname = lustre_cfg_string(lcfg, 1);
1627 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1628 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1629
1630 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1631 cfg->cfg_last_idx);
1632
1633 /* We're only called through here on the initial mount */
1634 rc = config_log_add(obd, logname, cfg, sb);
1635 if (rc)
1636 break;
1637 cld = config_log_find(logname, cfg);
1638 if (cld == NULL) {
1639 rc = -ENOENT;
1640 break;
1641 }
1642
1643 /* COMPAT_146 */
1644 /* FIXME only set this for old logs! Right now this forces
1645 us to always skip the "inside markers" check */
1646 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1647
1648 rc = mgc_process_log(obd, cld);
1649 if (rc == 0 && cld->cld_recover != NULL) {
1650 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1651 imp_connect_data, IMP_RECOV)) {
1652 rc = mgc_process_log(obd, cld->cld_recover);
1653 } else {
1654 struct config_llog_data *cir = cld->cld_recover;
1655
1656 cld->cld_recover = NULL;
1657 config_log_put(cir);
1658 }
1659 if (rc)
1660 CERROR("Cannot process recover llog %d\n", rc);
1661 }
1662
1663 if (rc == 0 && cld->cld_params != NULL) {
1664 rc = mgc_process_log(obd, cld->cld_params);
1665 if (rc == -ENOENT) {
1666 CDEBUG(D_MGC,
1667 "There is no params config file yet\n");
1668 rc = 0;
1669 }
1670 /* params log is optional */
1671 if (rc)
1672 CERROR(
1673 "%s: can't process params llog: rc = %d\n",
1674 obd->obd_name, rc);
1675 }
1676 config_log_put(cld);
1677
1678 break;
1679 }
1680 case LCFG_LOG_END: {
1681 logname = lustre_cfg_string(lcfg, 1);
1682
1683 if (lcfg->lcfg_bufcount >= 2)
1684 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1685 lcfg, 2);
1686 rc = config_log_end(logname, cfg);
1687 break;
1688 }
1689 default: {
1690 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1691 rc = -EINVAL;
1692 goto out;
1693
1694 }
1695 }
1696 out:
1697 return rc;
1698 }
1699
1700 static struct obd_ops mgc_obd_ops = {
1701 .o_owner = THIS_MODULE,
1702 .o_setup = mgc_setup,
1703 .o_precleanup = mgc_precleanup,
1704 .o_cleanup = mgc_cleanup,
1705 .o_add_conn = client_import_add_conn,
1706 .o_del_conn = client_import_del_conn,
1707 .o_connect = client_connect_import,
1708 .o_disconnect = client_disconnect_export,
1709 /* .o_enqueue = mgc_enqueue, */
1710 /* .o_iocontrol = mgc_iocontrol, */
1711 .o_set_info_async = mgc_set_info_async,
1712 .o_get_info = mgc_get_info,
1713 .o_import_event = mgc_import_event,
1714 .o_process_config = mgc_process_config,
1715 };
1716
mgc_init(void)1717 static int __init mgc_init(void)
1718 {
1719 return class_register_type(&mgc_obd_ops, NULL,
1720 LUSTRE_MGC_NAME, NULL);
1721 }
1722
mgc_exit(void)1723 static void /*__exit*/ mgc_exit(void)
1724 {
1725 class_unregister_type(LUSTRE_MGC_NAME);
1726 }
1727
1728 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1729 MODULE_DESCRIPTION("Lustre Management Client");
1730 MODULE_LICENSE("GPL");
1731
1732 module_init(mgc_init);
1733 module_exit(mgc_exit);
1734