projects / deliverable / linux.git / blob
? search:


f57d435e3a6b599159542fd86292458ccc9c9e08
[deliverable/linux.git] / drivers / staging / lustre / 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.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
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
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
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 */
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 */
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
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
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
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 lustre_sb_info *lsi = s2lsi(sb);
252 struct config_llog_data *cld;
253 char logname[32];
254
255 if (IS_OST(lsi))
256 return NULL;
257
258 /* for osp-on-ost, see lustre_start_osp() */
259 if (IS_MDT(lsi) && lcfg.cfg_instance)
260 return NULL;
261
262 /* we have to use different llog for clients and mdts for cmd
263 * where only clients are notified if one of cmd server restarts */
264 LASSERT(strlen(fsname) < sizeof(logname) / 2);
265 strcpy(logname, fsname);
266 LASSERT(lcfg.cfg_instance);
267 strcat(logname, "-cliir");
268
269 cld = do_config_log_add(obd, logname, CONFIG_T_RECOVER, &lcfg, sb);
270 return cld;
271 }
272
273 static struct config_llog_data *config_params_log_add(struct obd_device *obd,
274 struct config_llog_instance *cfg, struct super_block *sb)
275 {
276 struct config_llog_instance lcfg = *cfg;
277 struct config_llog_data *cld;
278
279 lcfg.cfg_instance = sb;
280
281 cld = do_config_log_add(obd, PARAMS_FILENAME, CONFIG_T_PARAMS,
282 &lcfg, sb);
283
284 return cld;
285 }
286
287 /** Add this log to the list of active logs watched by an MGC.
288 * Active means we're watching for updates.
289 * We have one active log per "mount" - client instance or servername.
290 * Each instance may be at a different point in the log.
291 */
292 static int config_log_add(struct obd_device *obd, char *logname,
293 struct config_llog_instance *cfg,
294 struct super_block *sb)
295 {
296 struct lustre_sb_info *lsi = s2lsi(sb);
297 struct config_llog_data *cld;
298 struct config_llog_data *sptlrpc_cld;
299 struct config_llog_data *params_cld;
300 char seclogname[32];
301 char *ptr;
302 int rc;
303
304 CDEBUG(D_MGC, "adding config log %s:%p\n", logname, cfg->cfg_instance);
305
306 /*
307 * for each regular log, the depended sptlrpc log name is
308 * <fsname>-sptlrpc. multiple regular logs may share one sptlrpc log.
309 */
310 ptr = strrchr(logname, '-');
311 if (ptr == NULL || ptr - logname > 8) {
312 CERROR("logname %s is too long\n", logname);
313 return -EINVAL;
314 }
315
316 memcpy(seclogname, logname, ptr - logname);
317 strcpy(seclogname + (ptr - logname), "-sptlrpc");
318
319 sptlrpc_cld = config_log_find(seclogname, NULL);
320 if (sptlrpc_cld == NULL) {
321 sptlrpc_cld = do_config_log_add(obd, seclogname,
322 CONFIG_T_SPTLRPC, NULL, NULL);
323 if (IS_ERR(sptlrpc_cld)) {
324 CERROR("can't create sptlrpc log: %s\n", seclogname);
325 rc = PTR_ERR(sptlrpc_cld);
326 goto out_err;
327 }
328 }
329 params_cld = config_params_log_add(obd, cfg, sb);
330 if (IS_ERR(params_cld)) {
331 rc = PTR_ERR(params_cld);
332 CERROR("%s: can't create params log: rc = %d\n",
333 obd->obd_name, rc);
334 goto out_err1;
335 }
336
337 cld = do_config_log_add(obd, logname, CONFIG_T_CONFIG, cfg, sb);
338 if (IS_ERR(cld)) {
339 CERROR("can't create log: %s\n", logname);
340 rc = PTR_ERR(cld);
341 goto out_err2;
342 }
343
344 cld->cld_sptlrpc = sptlrpc_cld;
345 cld->cld_params = params_cld;
346
347 LASSERT(lsi->lsi_lmd);
348 if (!(lsi->lsi_lmd->lmd_flags & LMD_FLG_NOIR)) {
349 struct config_llog_data *recover_cld;
350 *strrchr(seclogname, '-') = 0;
351 recover_cld = config_recover_log_add(obd, seclogname, cfg, sb);
352 if (IS_ERR(recover_cld)) {
353 rc = PTR_ERR(recover_cld);
354 goto out_err3;
355 }
356 cld->cld_recover = recover_cld;
357 }
358
359 return 0;
360
361 out_err3:
362 config_log_put(cld);
363
364 out_err2:
365 config_log_put(params_cld);
366
367 out_err1:
368 config_log_put(sptlrpc_cld);
369
370 out_err:
371 return rc;
372 }
373
374 DEFINE_MUTEX(llog_process_lock);
375
376 /** Stop watching for updates on this log.
377 */
378 static int config_log_end(char *logname, struct config_llog_instance *cfg)
379 {
380 struct config_llog_data *cld;
381 struct config_llog_data *cld_sptlrpc = NULL;
382 struct config_llog_data *cld_params = NULL;
383 struct config_llog_data *cld_recover = NULL;
384 int rc = 0;
385
386 cld = config_log_find(logname, cfg);
387 if (cld == NULL)
388 return -ENOENT;
389
390 mutex_lock(&cld->cld_lock);
391 /*
392 * if cld_stopping is set, it means we didn't start the log thus
393 * not owning the start ref. this can happen after previous umount:
394 * the cld still hanging there waiting for lock cancel, and we
395 * remount again but failed in the middle and call log_end without
396 * calling start_log.
397 */
398 if (unlikely(cld->cld_stopping)) {
399 mutex_unlock(&cld->cld_lock);
400 /* drop the ref from the find */
401 config_log_put(cld);
402 return rc;
403 }
404
405 cld->cld_stopping = 1;
406
407 cld_recover = cld->cld_recover;
408 cld->cld_recover = NULL;
409 mutex_unlock(&cld->cld_lock);
410
411 if (cld_recover) {
412 mutex_lock(&cld_recover->cld_lock);
413 cld_recover->cld_stopping = 1;
414 mutex_unlock(&cld_recover->cld_lock);
415 config_log_put(cld_recover);
416 }
417
418 spin_lock(&config_list_lock);
419 cld_sptlrpc = cld->cld_sptlrpc;
420 cld->cld_sptlrpc = NULL;
421 cld_params = cld->cld_params;
422 cld->cld_params = NULL;
423 spin_unlock(&config_list_lock);
424
425 if (cld_sptlrpc)
426 config_log_put(cld_sptlrpc);
427
428 if (cld_params) {
429 mutex_lock(&cld_params->cld_lock);
430 cld_params->cld_stopping = 1;
431 mutex_unlock(&cld_params->cld_lock);
432 config_log_put(cld_params);
433 }
434
435 /* drop the ref from the find */
436 config_log_put(cld);
437 /* drop the start ref */
438 config_log_put(cld);
439
440 CDEBUG(D_MGC, "end config log %s (%d)\n", logname ? logname : "client",
441 rc);
442 return rc;
443 }
444
445 int lprocfs_mgc_rd_ir_state(struct seq_file *m, void *data)
446 {
447 struct obd_device *obd = data;
448 struct obd_import *imp;
449 struct obd_connect_data *ocd;
450 struct config_llog_data *cld;
451
452 LPROCFS_CLIMP_CHECK(obd);
453 imp = obd->u.cli.cl_import;
454 ocd = &imp->imp_connect_data;
455
456 seq_printf(m, "imperative_recovery: %s\n",
457 OCD_HAS_FLAG(ocd, IMP_RECOV) ? "ENABLED" : "DISABLED");
458 seq_printf(m, "client_state:\n");
459
460 spin_lock(&config_list_lock);
461 list_for_each_entry(cld, &config_llog_list, cld_list_chain) {
462 if (cld->cld_recover == NULL)
463 continue;
464 seq_printf(m, " - { client: %s, nidtbl_version: %u }\n",
465 cld->cld_logname,
466 cld->cld_recover->cld_cfg.cfg_last_idx);
467 }
468 spin_unlock(&config_list_lock);
469
470 LPROCFS_CLIMP_EXIT(obd);
471 return 0;
472 }
473
474 /* reenqueue any lost locks */
475 #define RQ_RUNNING 0x1
476 #define RQ_NOW 0x2
477 #define RQ_LATER 0x4
478 #define RQ_STOP 0x8
479 #define RQ_PRECLEANUP 0x10
480 static int rq_state;
481 static wait_queue_head_t rq_waitq;
482 static DECLARE_COMPLETION(rq_exit);
483 static DECLARE_COMPLETION(rq_start);
484
485 static void do_requeue(struct config_llog_data *cld)
486 {
487 LASSERT(atomic_read(&cld->cld_refcount) > 0);
488
489 /* Do not run mgc_process_log on a disconnected export or an
490 export which is being disconnected. Take the client
491 semaphore to make the check non-racy. */
492 down_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
493 if (cld->cld_mgcexp->exp_obd->u.cli.cl_conn_count != 0) {
494 CDEBUG(D_MGC, "updating log %s\n", cld->cld_logname);
495 mgc_process_log(cld->cld_mgcexp->exp_obd, cld);
496 } else {
497 CDEBUG(D_MGC, "disconnecting, won't update log %s\n",
498 cld->cld_logname);
499 }
500 up_read(&cld->cld_mgcexp->exp_obd->u.cli.cl_sem);
501 }
502
503 /* this timeout represents how many seconds MGC should wait before
504 * requeue config and recover lock to the MGS. We need to randomize this
505 * in order to not flood the MGS.
506 */
507 #define MGC_TIMEOUT_MIN_SECONDS 5
508 #define MGC_TIMEOUT_RAND_CENTISEC 0x1ff /* ~500 */
509
510 static int mgc_requeue_thread(void *data)
511 {
512 bool first = true;
513
514 CDEBUG(D_MGC, "Starting requeue thread\n");
515
516 /* Keep trying failed locks periodically */
517 spin_lock(&config_list_lock);
518 rq_state |= RQ_RUNNING;
519 while (1) {
520 struct l_wait_info lwi;
521 struct config_llog_data *cld, *cld_prev;
522 int rand = cfs_rand() & MGC_TIMEOUT_RAND_CENTISEC;
523 int stopped = !!(rq_state & RQ_STOP);
524 int to;
525
526 /* Any new or requeued lostlocks will change the state */
527 rq_state &= ~(RQ_NOW | RQ_LATER);
528 spin_unlock(&config_list_lock);
529
530 if (first) {
531 first = false;
532 complete(&rq_start);
533 }
534
535 /* Always wait a few seconds to allow the server who
536 caused the lock revocation to finish its setup, plus some
537 random so everyone doesn't try to reconnect at once. */
538 to = MGC_TIMEOUT_MIN_SECONDS * HZ;
539 to += rand * HZ / 100; /* rand is centi-seconds */
540 lwi = LWI_TIMEOUT(to, NULL, NULL);
541 l_wait_event(rq_waitq, rq_state & (RQ_STOP | RQ_PRECLEANUP),
542 &lwi);
543
544 /*
545 * iterate & processing through the list. for each cld, process
546 * its depending sptlrpc cld firstly (if any) and then itself.
547 *
548 * it's guaranteed any item in the list must have
549 * reference > 0; and if cld_lostlock is set, at
550 * least one reference is taken by the previous enqueue.
551 */
552 cld_prev = NULL;
553
554 spin_lock(&config_list_lock);
555 rq_state &= ~RQ_PRECLEANUP;
556 list_for_each_entry(cld, &config_llog_list,
557 cld_list_chain) {
558 if (!cld->cld_lostlock)
559 continue;
560
561 spin_unlock(&config_list_lock);
562
563 LASSERT(atomic_read(&cld->cld_refcount) > 0);
564
565 /* Whether we enqueued again or not in mgc_process_log,
566 * we're done with the ref from the old enqueue */
567 if (cld_prev)
568 config_log_put(cld_prev);
569 cld_prev = cld;
570
571 cld->cld_lostlock = 0;
572 if (likely(!stopped))
573 do_requeue(cld);
574
575 spin_lock(&config_list_lock);
576 }
577 spin_unlock(&config_list_lock);
578 if (cld_prev)
579 config_log_put(cld_prev);
580
581 /* break after scanning the list so that we can drop
582 * refcount to losing lock clds */
583 if (unlikely(stopped)) {
584 spin_lock(&config_list_lock);
585 break;
586 }
587
588 /* Wait a bit to see if anyone else needs a requeue */
589 lwi = (struct l_wait_info) { 0 };
590 l_wait_event(rq_waitq, rq_state & (RQ_NOW | RQ_STOP),
591 &lwi);
592 spin_lock(&config_list_lock);
593 }
594 /* spinlock and while guarantee RQ_NOW and RQ_LATER are not set */
595 rq_state &= ~RQ_RUNNING;
596 spin_unlock(&config_list_lock);
597
598 complete(&rq_exit);
599
600 CDEBUG(D_MGC, "Ending requeue thread\n");
601 return 0;
602 }
603
604 /* Add a cld to the list to requeue. Start the requeue thread if needed.
605 We are responsible for dropping the config log reference from here on out. */
606 static void mgc_requeue_add(struct config_llog_data *cld)
607 {
608 CDEBUG(D_INFO, "log %s: requeue (r=%d sp=%d st=%x)\n",
609 cld->cld_logname, atomic_read(&cld->cld_refcount),
610 cld->cld_stopping, rq_state);
611 LASSERT(atomic_read(&cld->cld_refcount) > 0);
612
613 mutex_lock(&cld->cld_lock);
614 if (cld->cld_stopping || cld->cld_lostlock) {
615 mutex_unlock(&cld->cld_lock);
616 return;
617 }
618 /* this refcount will be released in mgc_requeue_thread. */
619 config_log_get(cld);
620 cld->cld_lostlock = 1;
621 mutex_unlock(&cld->cld_lock);
622
623 /* Hold lock for rq_state */
624 spin_lock(&config_list_lock);
625 if (rq_state & RQ_STOP) {
626 spin_unlock(&config_list_lock);
627 cld->cld_lostlock = 0;
628 config_log_put(cld);
629 } else {
630 rq_state |= RQ_NOW;
631 spin_unlock(&config_list_lock);
632 wake_up(&rq_waitq);
633 }
634 }
635
636 static int mgc_llog_init(const struct lu_env *env, struct obd_device *obd)
637 {
638 struct llog_ctxt *ctxt;
639 int rc;
640
641 /* setup only remote ctxt, the local disk context is switched per each
642 * filesystem during mgc_fs_setup() */
643 rc = llog_setup(env, obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, obd,
644 &llog_client_ops);
645 if (rc)
646 return rc;
647
648 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
649 LASSERT(ctxt);
650
651 llog_initiator_connect(ctxt);
652 llog_ctxt_put(ctxt);
653
654 return 0;
655 }
656
657 static int mgc_llog_fini(const struct lu_env *env, struct obd_device *obd)
658 {
659 struct llog_ctxt *ctxt;
660
661 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
662 if (ctxt)
663 llog_cleanup(env, ctxt);
664
665 return 0;
666 }
667
668 static atomic_t mgc_count = ATOMIC_INIT(0);
669 static int mgc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
670 {
671 int rc = 0;
672 int temp;
673
674 switch (stage) {
675 case OBD_CLEANUP_EARLY:
676 break;
677 case OBD_CLEANUP_EXPORTS:
678 if (atomic_dec_and_test(&mgc_count)) {
679 LASSERT(rq_state & RQ_RUNNING);
680 /* stop requeue thread */
681 temp = RQ_STOP;
682 } else {
683 /* wakeup requeue thread to clean our cld */
684 temp = RQ_NOW | RQ_PRECLEANUP;
685 }
686 spin_lock(&config_list_lock);
687 rq_state |= temp;
688 spin_unlock(&config_list_lock);
689 wake_up(&rq_waitq);
690 if (temp & RQ_STOP)
691 wait_for_completion(&rq_exit);
692 obd_cleanup_client_import(obd);
693 rc = mgc_llog_fini(NULL, obd);
694 if (rc != 0)
695 CERROR("failed to cleanup llogging subsystems\n");
696 break;
697 }
698 return rc;
699 }
700
701 static int mgc_cleanup(struct obd_device *obd)
702 {
703 /* COMPAT_146 - old config logs may have added profiles we don't
704 know about */
705 if (obd->obd_type->typ_refcnt <= 1)
706 /* Only for the last mgc */
707 class_del_profiles();
708
709 lprocfs_obd_cleanup(obd);
710 ptlrpcd_decref();
711
712 return client_obd_cleanup(obd);
713 }
714
715 static int mgc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
716 {
717 struct lprocfs_static_vars lvars = { NULL };
718 int rc;
719
720 ptlrpcd_addref();
721
722 rc = client_obd_setup(obd, lcfg);
723 if (rc)
724 goto err_decref;
725
726 rc = mgc_llog_init(NULL, obd);
727 if (rc) {
728 CERROR("failed to setup llogging subsystems\n");
729 goto err_cleanup;
730 }
731
732 lprocfs_mgc_init_vars(&lvars);
733 lprocfs_obd_setup(obd, lvars.obd_vars, lvars.sysfs_vars);
734 sptlrpc_lprocfs_cliobd_attach(obd);
735
736 if (atomic_inc_return(&mgc_count) == 1) {
737 rq_state = 0;
738 init_waitqueue_head(&rq_waitq);
739
740 /* start requeue thread */
741 rc = PTR_ERR(kthread_run(mgc_requeue_thread, NULL,
742 "ll_cfg_requeue"));
743 if (IS_ERR_VALUE(rc)) {
744 CERROR("%s: Cannot start requeue thread (%d),no more log updates!\n",
745 obd->obd_name, rc);
746 goto err_cleanup;
747 }
748 /* rc is the task_struct pointer of mgc_requeue_thread. */
749 rc = 0;
750 wait_for_completion(&rq_start);
751 }
752
753 return rc;
754
755 err_cleanup:
756 client_obd_cleanup(obd);
757 err_decref:
758 ptlrpcd_decref();
759 return rc;
760 }
761
762 /* based on ll_mdc_blocking_ast */
763 static int mgc_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
764 void *data, int flag)
765 {
766 struct lustre_handle lockh;
767 struct config_llog_data *cld = (struct config_llog_data *)data;
768 int rc = 0;
769
770 switch (flag) {
771 case LDLM_CB_BLOCKING:
772 /* mgs wants the lock, give it up... */
773 LDLM_DEBUG(lock, "MGC blocking CB");
774 ldlm_lock2handle(lock, &lockh);
775 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
776 break;
777 case LDLM_CB_CANCELING:
778 /* We've given up the lock, prepare ourselves to update. */
779 LDLM_DEBUG(lock, "MGC cancel CB");
780
781 CDEBUG(D_MGC, "Lock res "DLDLMRES" (%.8s)\n",
782 PLDLMRES(lock->l_resource),
783 (char *)&lock->l_resource->lr_name.name[0]);
784
785 if (!cld) {
786 CDEBUG(D_INFO, "missing data, won't requeue\n");
787 break;
788 }
789
790 /* held at mgc_process_log(). */
791 LASSERT(atomic_read(&cld->cld_refcount) > 0);
792 /* Are we done with this log? */
793 if (cld->cld_stopping) {
794 CDEBUG(D_MGC, "log %s: stopping, won't requeue\n",
795 cld->cld_logname);
796 config_log_put(cld);
797 break;
798 }
799 /* Make sure not to re-enqueue when the mgc is stopping
800 (we get called from client_disconnect_export) */
801 if (!lock->l_conn_export ||
802 !lock->l_conn_export->exp_obd->u.cli.cl_conn_count) {
803 CDEBUG(D_MGC, "log %.8s: disconnecting, won't requeue\n",
804 cld->cld_logname);
805 config_log_put(cld);
806 break;
807 }
808
809 /* Re-enqueue now */
810 mgc_requeue_add(cld);
811 config_log_put(cld);
812 break;
813 default:
814 LBUG();
815 }
816
817 return rc;
818 }
819
820 /* Not sure where this should go... */
821 /* This is the timeout value for MGS_CONNECT request plus a ping interval, such
822 * that we can have a chance to try the secondary MGS if any. */
823 #define MGC_ENQUEUE_LIMIT (INITIAL_CONNECT_TIMEOUT + (AT_OFF ? 0 : at_min) \
824 + PING_INTERVAL)
825 #define MGC_TARGET_REG_LIMIT 10
826 #define MGC_SEND_PARAM_LIMIT 10
827
828 /* Send parameter to MGS*/
829 static int mgc_set_mgs_param(struct obd_export *exp,
830 struct mgs_send_param *msp)
831 {
832 struct ptlrpc_request *req;
833 struct mgs_send_param *req_msp, *rep_msp;
834 int rc;
835
836 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
837 &RQF_MGS_SET_INFO, LUSTRE_MGS_VERSION,
838 MGS_SET_INFO);
839 if (!req)
840 return -ENOMEM;
841
842 req_msp = req_capsule_client_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
843 if (!req_msp) {
844 ptlrpc_req_finished(req);
845 return -ENOMEM;
846 }
847
848 memcpy(req_msp, msp, sizeof(*req_msp));
849 ptlrpc_request_set_replen(req);
850
851 /* Limit how long we will wait for the enqueue to complete */
852 req->rq_delay_limit = MGC_SEND_PARAM_LIMIT;
853 rc = ptlrpc_queue_wait(req);
854 if (!rc) {
855 rep_msp = req_capsule_server_get(&req->rq_pill, &RMF_MGS_SEND_PARAM);
856 memcpy(msp, rep_msp, sizeof(*rep_msp));
857 }
858
859 ptlrpc_req_finished(req);
860
861 return rc;
862 }
863
864 /* Take a config lock so we can get cancel notifications */
865 static int mgc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
866 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
867 __u64 *flags, void *bl_cb, void *cp_cb, void *gl_cb,
868 void *data, __u32 lvb_len, void *lvb_swabber,
869 struct lustre_handle *lockh)
870 {
871 struct config_llog_data *cld = (struct config_llog_data *)data;
872 struct ldlm_enqueue_info einfo = {
873 .ei_type = type,
874 .ei_mode = mode,
875 .ei_cb_bl = mgc_blocking_ast,
876 .ei_cb_cp = ldlm_completion_ast,
877 };
878 struct ptlrpc_request *req;
879 int short_limit = cld_is_sptlrpc(cld);
880 int rc;
881
882 CDEBUG(D_MGC, "Enqueue for %s (res %#llx)\n", cld->cld_logname,
883 cld->cld_resid.name[0]);
884
885 /* We need a callback for every lockholder, so don't try to
886 ldlm_lock_match (see rev 1.1.2.11.2.47) */
887 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
888 &RQF_LDLM_ENQUEUE, LUSTRE_DLM_VERSION,
889 LDLM_ENQUEUE);
890 if (req == NULL)
891 return -ENOMEM;
892
893 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, 0);
894 ptlrpc_request_set_replen(req);
895
896 /* Limit how long we will wait for the enqueue to complete */
897 req->rq_delay_limit = short_limit ? 5 : MGC_ENQUEUE_LIMIT;
898 rc = ldlm_cli_enqueue(exp, &req, &einfo, &cld->cld_resid, NULL, flags,
899 NULL, 0, LVB_T_NONE, lockh, 0);
900 /* A failed enqueue should still call the mgc_blocking_ast,
901 where it will be requeued if needed ("grant failed"). */
902 ptlrpc_req_finished(req);
903 return rc;
904 }
905
906 static void mgc_notify_active(struct obd_device *unused)
907 {
908 /* wakeup mgc_requeue_thread to requeue mgc lock */
909 spin_lock(&config_list_lock);
910 rq_state |= RQ_NOW;
911 spin_unlock(&config_list_lock);
912 wake_up(&rq_waitq);
913
914 /* TODO: Help the MGS rebuild nidtbl. -jay */
915 }
916
917 /* Send target_reg message to MGS */
918 static int mgc_target_register(struct obd_export *exp,
919 struct mgs_target_info *mti)
920 {
921 struct ptlrpc_request *req;
922 struct mgs_target_info *req_mti, *rep_mti;
923 int rc;
924
925 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
926 &RQF_MGS_TARGET_REG, LUSTRE_MGS_VERSION,
927 MGS_TARGET_REG);
928 if (req == NULL)
929 return -ENOMEM;
930
931 req_mti = req_capsule_client_get(&req->rq_pill, &RMF_MGS_TARGET_INFO);
932 if (!req_mti) {
933 ptlrpc_req_finished(req);
934 return -ENOMEM;
935 }
936
937 memcpy(req_mti, mti, sizeof(*req_mti));
938 ptlrpc_request_set_replen(req);
939 CDEBUG(D_MGC, "register %s\n", mti->mti_svname);
940 /* Limit how long we will wait for the enqueue to complete */
941 req->rq_delay_limit = MGC_TARGET_REG_LIMIT;
942
943 rc = ptlrpc_queue_wait(req);
944 if (!rc) {
945 rep_mti = req_capsule_server_get(&req->rq_pill,
946 &RMF_MGS_TARGET_INFO);
947 memcpy(mti, rep_mti, sizeof(*rep_mti));
948 CDEBUG(D_MGC, "register %s got index = %d\n",
949 mti->mti_svname, mti->mti_stripe_index);
950 }
951 ptlrpc_req_finished(req);
952
953 return rc;
954 }
955
956 static int mgc_set_info_async(const struct lu_env *env, struct obd_export *exp,
957 u32 keylen, void *key, u32 vallen,
958 void *val, struct ptlrpc_request_set *set)
959 {
960 int rc = -EINVAL;
961
962 /* Turn off initial_recov after we try all backup servers once */
963 if (KEY_IS(KEY_INIT_RECOV_BACKUP)) {
964 struct obd_import *imp = class_exp2cliimp(exp);
965 int value;
966 if (vallen != sizeof(int))
967 return -EINVAL;
968 value = *(int *)val;
969 CDEBUG(D_MGC, "InitRecov %s %d/d%d:i%d:r%d:or%d:%s\n",
970 imp->imp_obd->obd_name, value,
971 imp->imp_deactive, imp->imp_invalid,
972 imp->imp_replayable, imp->imp_obd->obd_replayable,
973 ptlrpc_import_state_name(imp->imp_state));
974 /* Resurrect if we previously died */
975 if ((imp->imp_state != LUSTRE_IMP_FULL &&
976 imp->imp_state != LUSTRE_IMP_NEW) || value > 1)
977 ptlrpc_reconnect_import(imp);
978 return 0;
979 }
980 if (KEY_IS(KEY_SET_INFO)) {
981 struct mgs_send_param *msp;
982
983 msp = (struct mgs_send_param *)val;
984 rc = mgc_set_mgs_param(exp, msp);
985 return rc;
986 }
987 if (KEY_IS(KEY_MGSSEC)) {
988 struct client_obd *cli = &exp->exp_obd->u.cli;
989 struct sptlrpc_flavor flvr;
990
991 /*
992 * empty string means using current flavor, if which haven't
993 * been set yet, set it as null.
994 *
995 * if flavor has been set previously, check the asking flavor
996 * must match the existing one.
997 */
998 if (vallen == 0) {
999 if (cli->cl_flvr_mgc.sf_rpc != SPTLRPC_FLVR_INVALID)
1000 return 0;
1001 val = "null";
1002 vallen = 4;
1003 }
1004
1005 rc = sptlrpc_parse_flavor(val, &flvr);
1006 if (rc) {
1007 CERROR("invalid sptlrpc flavor %s to MGS\n",
1008 (char *) val);
1009 return rc;
1010 }
1011
1012 /*
1013 * caller already hold a mutex
1014 */
1015 if (cli->cl_flvr_mgc.sf_rpc == SPTLRPC_FLVR_INVALID) {
1016 cli->cl_flvr_mgc = flvr;
1017 } else if (memcmp(&cli->cl_flvr_mgc, &flvr,
1018 sizeof(flvr)) != 0) {
1019 char str[20];
1020
1021 sptlrpc_flavor2name(&cli->cl_flvr_mgc,
1022 str, sizeof(str));
1023 LCONSOLE_ERROR("asking sptlrpc flavor %s to MGS but currently %s is in use\n",
1024 (char *) val, str);
1025 rc = -EPERM;
1026 }
1027 return rc;
1028 }
1029
1030 return rc;
1031 }
1032
1033 static int mgc_get_info(const struct lu_env *env, struct obd_export *exp,
1034 __u32 keylen, void *key, __u32 *vallen, void *val,
1035 struct lov_stripe_md *unused)
1036 {
1037 int rc = -EINVAL;
1038
1039 if (KEY_IS(KEY_CONN_DATA)) {
1040 struct obd_import *imp = class_exp2cliimp(exp);
1041 struct obd_connect_data *data = val;
1042
1043 if (*vallen == sizeof(*data)) {
1044 *data = imp->imp_connect_data;
1045 rc = 0;
1046 }
1047 }
1048
1049 return rc;
1050 }
1051
1052 static int mgc_import_event(struct obd_device *obd,
1053 struct obd_import *imp,
1054 enum obd_import_event event)
1055 {
1056 LASSERT(imp->imp_obd == obd);
1057 CDEBUG(D_MGC, "import event %#x\n", event);
1058
1059 switch (event) {
1060 case IMP_EVENT_DISCON:
1061 /* MGC imports should not wait for recovery */
1062 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1063 ptlrpc_pinger_ir_down();
1064 break;
1065 case IMP_EVENT_INACTIVE:
1066 break;
1067 case IMP_EVENT_INVALIDATE: {
1068 struct ldlm_namespace *ns = obd->obd_namespace;
1069 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
1070 break;
1071 }
1072 case IMP_EVENT_ACTIVE:
1073 CDEBUG(D_INFO, "%s: Reactivating import\n", obd->obd_name);
1074 /* Clearing obd_no_recov allows us to continue pinging */
1075 obd->obd_no_recov = 0;
1076 mgc_notify_active(obd);
1077 if (OCD_HAS_FLAG(&imp->imp_connect_data, IMP_RECOV))
1078 ptlrpc_pinger_ir_up();
1079 break;
1080 case IMP_EVENT_OCD:
1081 break;
1082 case IMP_EVENT_DEACTIVATE:
1083 case IMP_EVENT_ACTIVATE:
1084 break;
1085 default:
1086 CERROR("Unknown import event %#x\n", event);
1087 LBUG();
1088 }
1089 return 0;
1090 }
1091
1092 enum {
1093 CONFIG_READ_NRPAGES_INIT = 1 << (20 - PAGE_CACHE_SHIFT),
1094 CONFIG_READ_NRPAGES = 4
1095 };
1096
1097 static int mgc_apply_recover_logs(struct obd_device *mgc,
1098 struct config_llog_data *cld,
1099 __u64 max_version,
1100 void *data, int datalen, bool mne_swab)
1101 {
1102 struct config_llog_instance *cfg = &cld->cld_cfg;
1103 struct mgs_nidtbl_entry *entry;
1104 struct lustre_cfg *lcfg;
1105 struct lustre_cfg_bufs bufs;
1106 u64 prev_version = 0;
1107 char *inst;
1108 char *buf;
1109 int bufsz;
1110 int pos;
1111 int rc = 0;
1112 int off = 0;
1113
1114 LASSERT(cfg->cfg_instance != NULL);
1115 LASSERT(cfg->cfg_sb == cfg->cfg_instance);
1116
1117 inst = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1118 if (!inst)
1119 return -ENOMEM;
1120
1121 pos = snprintf(inst, PAGE_CACHE_SIZE, "%p", cfg->cfg_instance);
1122 if (pos >= PAGE_CACHE_SIZE) {
1123 kfree(inst);
1124 return -E2BIG;
1125 }
1126
1127 ++pos;
1128 buf = inst + pos;
1129 bufsz = PAGE_CACHE_SIZE - pos;
1130
1131 while (datalen > 0) {
1132 int entry_len = sizeof(*entry);
1133 int is_ost;
1134 struct obd_device *obd;
1135 char *obdname;
1136 char *cname;
1137 char *params;
1138 char *uuid;
1139
1140 rc = -EINVAL;
1141 if (datalen < sizeof(*entry))
1142 break;
1143
1144 entry = (typeof(entry))(data + off);
1145
1146 /* sanity check */
1147 if (entry->mne_nid_type != 0) /* only support type 0 for ipv4 */
1148 break;
1149 if (entry->mne_nid_count == 0) /* at least one nid entry */
1150 break;
1151 if (entry->mne_nid_size != sizeof(lnet_nid_t))
1152 break;
1153
1154 entry_len += entry->mne_nid_count * entry->mne_nid_size;
1155 if (datalen < entry_len) /* must have entry_len at least */
1156 break;
1157
1158 /* Keep this swab for normal mixed endian handling. LU-1644 */
1159 if (mne_swab)
1160 lustre_swab_mgs_nidtbl_entry(entry);
1161 if (entry->mne_length > PAGE_CACHE_SIZE) {
1162 CERROR("MNE too large (%u)\n", entry->mne_length);
1163 break;
1164 }
1165
1166 if (entry->mne_length < entry_len)
1167 break;
1168
1169 off += entry->mne_length;
1170 datalen -= entry->mne_length;
1171 if (datalen < 0)
1172 break;
1173
1174 if (entry->mne_version > max_version) {
1175 CERROR("entry index(%lld) is over max_index(%lld)\n",
1176 entry->mne_version, max_version);
1177 break;
1178 }
1179
1180 if (prev_version >= entry->mne_version) {
1181 CERROR("index unsorted, prev %lld, now %lld\n",
1182 prev_version, entry->mne_version);
1183 break;
1184 }
1185 prev_version = entry->mne_version;
1186
1187 /*
1188 * Write a string with format "nid::instance" to
1189 * lustre/<osc|mdc>/<target>-<osc|mdc>-<instance>/import.
1190 */
1191
1192 is_ost = entry->mne_type == LDD_F_SV_TYPE_OST;
1193 memset(buf, 0, bufsz);
1194 obdname = buf;
1195 pos = 0;
1196
1197 /* lustre-OST0001-osc-<instance #> */
1198 strcpy(obdname, cld->cld_logname);
1199 cname = strrchr(obdname, '-');
1200 if (cname == NULL) {
1201 CERROR("mgc %s: invalid logname %s\n",
1202 mgc->obd_name, obdname);
1203 break;
1204 }
1205
1206 pos = cname - obdname;
1207 obdname[pos] = 0;
1208 pos += sprintf(obdname + pos, "-%s%04x",
1209 is_ost ? "OST" : "MDT", entry->mne_index);
1210
1211 cname = is_ost ? "osc" : "mdc";
1212 pos += sprintf(obdname + pos, "-%s-%s", cname, inst);
1213 lustre_cfg_bufs_reset(&bufs, obdname);
1214
1215 /* find the obd by obdname */
1216 obd = class_name2obd(obdname);
1217 if (obd == NULL) {
1218 CDEBUG(D_INFO, "mgc %s: cannot find obdname %s\n",
1219 mgc->obd_name, obdname);
1220 rc = 0;
1221 /* this is a safe race, when the ost is starting up...*/
1222 continue;
1223 }
1224
1225 /* osc.import = "connection=<Conn UUID>::<target instance>" */
1226 ++pos;
1227 params = buf + pos;
1228 pos += sprintf(params, "%s.import=%s", cname, "connection=");
1229 uuid = buf + pos;
1230
1231 down_read(&obd->u.cli.cl_sem);
1232 if (obd->u.cli.cl_import == NULL) {
1233 /* client does not connect to the OST yet */
1234 up_read(&obd->u.cli.cl_sem);
1235 rc = 0;
1236 continue;
1237 }
1238
1239 /* TODO: iterate all nids to find one */
1240 /* find uuid by nid */
1241 rc = client_import_find_conn(obd->u.cli.cl_import,
1242 entry->u.nids[0],
1243 (struct obd_uuid *)uuid);
1244 up_read(&obd->u.cli.cl_sem);
1245 if (rc < 0) {
1246 CERROR("mgc: cannot find uuid by nid %s\n",
1247 libcfs_nid2str(entry->u.nids[0]));
1248 break;
1249 }
1250
1251 CDEBUG(D_INFO, "Find uuid %s by nid %s\n",
1252 uuid, libcfs_nid2str(entry->u.nids[0]));
1253
1254 pos += strlen(uuid);
1255 pos += sprintf(buf + pos, "::%u", entry->mne_instance);
1256 LASSERT(pos < bufsz);
1257
1258 lustre_cfg_bufs_set_string(&bufs, 1, params);
1259
1260 rc = -ENOMEM;
1261 lcfg = lustre_cfg_new(LCFG_PARAM, &bufs);
1262 if (lcfg == NULL) {
1263 CERROR("mgc: cannot allocate memory\n");
1264 break;
1265 }
1266
1267 CDEBUG(D_INFO, "ir apply logs %lld/%lld for %s -> %s\n",
1268 prev_version, max_version, obdname, params);
1269
1270 rc = class_process_config(lcfg);
1271 lustre_cfg_free(lcfg);
1272 if (rc)
1273 CDEBUG(D_INFO, "process config for %s error %d\n",
1274 obdname, rc);
1275
1276 /* continue, even one with error */
1277 }
1278
1279 kfree(inst);
1280 return rc;
1281 }
1282
1283 /**
1284 * This function is called if this client was notified for target restarting
1285 * by the MGS. A CONFIG_READ RPC is going to send to fetch recovery logs.
1286 */
1287 static int mgc_process_recover_log(struct obd_device *obd,
1288 struct config_llog_data *cld)
1289 {
1290 struct ptlrpc_request *req = NULL;
1291 struct config_llog_instance *cfg = &cld->cld_cfg;
1292 struct mgs_config_body *body;
1293 struct mgs_config_res *res;
1294 struct ptlrpc_bulk_desc *desc;
1295 struct page **pages;
1296 int nrpages;
1297 bool eof = true;
1298 bool mne_swab = false;
1299 int i;
1300 int ealen;
1301 int rc;
1302
1303 /* allocate buffer for bulk transfer.
1304 * if this is the first time for this mgs to read logs,
1305 * CONFIG_READ_NRPAGES_INIT will be used since it will read all logs
1306 * once; otherwise, it only reads increment of logs, this should be
1307 * small and CONFIG_READ_NRPAGES will be used.
1308 */
1309 nrpages = CONFIG_READ_NRPAGES;
1310 if (cfg->cfg_last_idx == 0) /* the first time */
1311 nrpages = CONFIG_READ_NRPAGES_INIT;
1312
1313 pages = kcalloc(nrpages, sizeof(*pages), GFP_NOFS);
1314 if (pages == NULL) {
1315 rc = -ENOMEM;
1316 goto out;
1317 }
1318
1319 for (i = 0; i < nrpages; i++) {
1320 pages[i] = alloc_page(GFP_IOFS);
1321 if (pages[i] == NULL) {
1322 rc = -ENOMEM;
1323 goto out;
1324 }
1325 }
1326
1327 again:
1328 LASSERT(cld_is_recover(cld));
1329 LASSERT(mutex_is_locked(&cld->cld_lock));
1330 req = ptlrpc_request_alloc(class_exp2cliimp(cld->cld_mgcexp),
1331 &RQF_MGS_CONFIG_READ);
1332 if (req == NULL) {
1333 rc = -ENOMEM;
1334 goto out;
1335 }
1336
1337 rc = ptlrpc_request_pack(req, LUSTRE_MGS_VERSION, MGS_CONFIG_READ);
1338 if (rc)
1339 goto out;
1340
1341 /* pack request */
1342 body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1343 LASSERT(body != NULL);
1344 LASSERT(sizeof(body->mcb_name) > strlen(cld->cld_logname));
1345 if (strlcpy(body->mcb_name, cld->cld_logname, sizeof(body->mcb_name))
1346 >= sizeof(body->mcb_name)) {
1347 rc = -E2BIG;
1348 goto out;
1349 }
1350 body->mcb_offset = cfg->cfg_last_idx + 1;
1351 body->mcb_type = cld->cld_type;
1352 body->mcb_bits = PAGE_CACHE_SHIFT;
1353 body->mcb_units = nrpages;
1354
1355 /* allocate bulk transfer descriptor */
1356 desc = ptlrpc_prep_bulk_imp(req, nrpages, 1, BULK_PUT_SINK,
1357 MGS_BULK_PORTAL);
1358 if (desc == NULL) {
1359 rc = -ENOMEM;
1360 goto out;
1361 }
1362
1363 for (i = 0; i < nrpages; i++)
1364 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
1365
1366 ptlrpc_request_set_replen(req);
1367 rc = ptlrpc_queue_wait(req);
1368 if (rc)
1369 goto out;
1370
1371 res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1372 if (res->mcr_size < res->mcr_offset) {
1373 rc = -EINVAL;
1374 goto out;
1375 }
1376
1377 /* always update the index even though it might have errors with
1378 * handling the recover logs */
1379 cfg->cfg_last_idx = res->mcr_offset;
1380 eof = res->mcr_offset == res->mcr_size;
1381
1382 CDEBUG(D_INFO, "Latest version %lld, more %d.\n",
1383 res->mcr_offset, eof == false);
1384
1385 ealen = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, 0);
1386 if (ealen < 0) {
1387 rc = ealen;
1388 goto out;
1389 }
1390
1391 if (ealen > nrpages << PAGE_CACHE_SHIFT) {
1392 rc = -EINVAL;
1393 goto out;
1394 }
1395
1396 if (ealen == 0) { /* no logs transferred */
1397 if (!eof)
1398 rc = -EINVAL;
1399 goto out;
1400 }
1401
1402 mne_swab = !!ptlrpc_rep_need_swab(req);
1403 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1404 /* This import flag means the server did an extra swab of IR MNE
1405 * records (fixed in LU-1252), reverse it here if needed. LU-1644 */
1406 if (unlikely(req->rq_import->imp_need_mne_swab))
1407 mne_swab = !mne_swab;
1408 #else
1409 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1410 #endif
1411
1412 for (i = 0; i < nrpages && ealen > 0; i++) {
1413 int rc2;
1414 void *ptr;
1415
1416 ptr = kmap(pages[i]);
1417 rc2 = mgc_apply_recover_logs(obd, cld, res->mcr_offset, ptr,
1418 min_t(int, ealen, PAGE_CACHE_SIZE),
1419 mne_swab);
1420 kunmap(pages[i]);
1421 if (rc2 < 0) {
1422 CWARN("Process recover log %s error %d\n",
1423 cld->cld_logname, rc2);
1424 break;
1425 }
1426
1427 ealen -= PAGE_CACHE_SIZE;
1428 }
1429
1430 out:
1431 if (req)
1432 ptlrpc_req_finished(req);
1433
1434 if (rc == 0 && !eof)
1435 goto again;
1436
1437 if (pages) {
1438 for (i = 0; i < nrpages; i++) {
1439 if (pages[i] == NULL)
1440 break;
1441 __free_page(pages[i]);
1442 }
1443 kfree(pages);
1444 }
1445 return rc;
1446 }
1447
1448 /* local_only means it cannot get remote llogs */
1449 static int mgc_process_cfg_log(struct obd_device *mgc,
1450 struct config_llog_data *cld, int local_only)
1451 {
1452 struct llog_ctxt *ctxt;
1453 struct lustre_sb_info *lsi = NULL;
1454 int rc = 0;
1455 bool sptlrpc_started = false;
1456 struct lu_env *env;
1457
1458 LASSERT(cld);
1459 LASSERT(mutex_is_locked(&cld->cld_lock));
1460
1461 /*
1462 * local copy of sptlrpc log is controlled elsewhere, don't try to
1463 * read it up here.
1464 */
1465 if (cld_is_sptlrpc(cld) && local_only)
1466 return 0;
1467
1468 if (cld->cld_cfg.cfg_sb)
1469 lsi = s2lsi(cld->cld_cfg.cfg_sb);
1470
1471 env = kzalloc(sizeof(*env), GFP_NOFS);
1472 if (!env)
1473 return -ENOMEM;
1474
1475 rc = lu_env_init(env, LCT_MG_THREAD);
1476 if (rc)
1477 goto out_free;
1478
1479 ctxt = llog_get_context(mgc, LLOG_CONFIG_REPL_CTXT);
1480 LASSERT(ctxt);
1481
1482 if (local_only) /* no local log at client side */ {
1483 rc = -EIO;
1484 goto out_pop;
1485 }
1486
1487 if (cld_is_sptlrpc(cld)) {
1488 sptlrpc_conf_log_update_begin(cld->cld_logname);
1489 sptlrpc_started = true;
1490 }
1491
1492 /* logname and instance info should be the same, so use our
1493 * copy of the instance for the update. The cfg_last_idx will
1494 * be updated here. */
1495 rc = class_config_parse_llog(env, ctxt, cld->cld_logname,
1496 &cld->cld_cfg);
1497
1498 out_pop:
1499 __llog_ctxt_put(env, ctxt);
1500
1501 /*
1502 * update settings on existing OBDs. doing it inside
1503 * of llog_process_lock so no device is attaching/detaching
1504 * in parallel.
1505 * the logname must be <fsname>-sptlrpc
1506 */
1507 if (sptlrpc_started) {
1508 LASSERT(cld_is_sptlrpc(cld));
1509 sptlrpc_conf_log_update_end(cld->cld_logname);
1510 class_notify_sptlrpc_conf(cld->cld_logname,
1511 strlen(cld->cld_logname) -
1512 strlen("-sptlrpc"));
1513 }
1514
1515 lu_env_fini(env);
1516 out_free:
1517 kfree(env);
1518 return rc;
1519 }
1520
1521 /** Get a config log from the MGS and process it.
1522 * This func is called for both clients and servers.
1523 * Copy the log locally before parsing it if appropriate (non-MGS server)
1524 */
1525 int mgc_process_log(struct obd_device *mgc, struct config_llog_data *cld)
1526 {
1527 struct lustre_handle lockh = { 0 };
1528 __u64 flags = LDLM_FL_NO_LRU;
1529 int rc = 0, rcl;
1530
1531 LASSERT(cld);
1532
1533 /* I don't want multiple processes running process_log at once --
1534 sounds like badness. It actually might be fine, as long as
1535 we're not trying to update from the same log
1536 simultaneously (in which case we should use a per-log sem.) */
1537 mutex_lock(&cld->cld_lock);
1538 if (cld->cld_stopping) {
1539 mutex_unlock(&cld->cld_lock);
1540 return 0;
1541 }
1542
1543 OBD_FAIL_TIMEOUT(OBD_FAIL_MGC_PAUSE_PROCESS_LOG, 20);
1544
1545 CDEBUG(D_MGC, "Process log %s:%p from %d\n", cld->cld_logname,
1546 cld->cld_cfg.cfg_instance, cld->cld_cfg.cfg_last_idx + 1);
1547
1548 /* Get the cfg lock on the llog */
1549 rcl = mgc_enqueue(mgc->u.cli.cl_mgc_mgsexp, NULL, LDLM_PLAIN, NULL,
1550 LCK_CR, &flags, NULL, NULL, NULL,
1551 cld, 0, NULL, &lockh);
1552 if (rcl == 0) {
1553 /* Get the cld, it will be released in mgc_blocking_ast. */
1554 config_log_get(cld);
1555 rc = ldlm_lock_set_data(&lockh, (void *)cld);
1556 LASSERT(rc == 0);
1557 } else {
1558 CDEBUG(D_MGC, "Can't get cfg lock: %d\n", rcl);
1559
1560 /* mark cld_lostlock so that it will requeue
1561 * after MGC becomes available. */
1562 cld->cld_lostlock = 1;
1563 /* Get extra reference, it will be put in requeue thread */
1564 config_log_get(cld);
1565 }
1566
1567
1568 if (cld_is_recover(cld)) {
1569 rc = 0; /* this is not a fatal error for recover log */
1570 if (rcl == 0)
1571 rc = mgc_process_recover_log(mgc, cld);
1572 } else {
1573 rc = mgc_process_cfg_log(mgc, cld, rcl != 0);
1574 }
1575
1576 CDEBUG(D_MGC, "%s: configuration from log '%s' %sed (%d).\n",
1577 mgc->obd_name, cld->cld_logname, rc ? "fail" : "succeed", rc);
1578
1579 mutex_unlock(&cld->cld_lock);
1580
1581 /* Now drop the lock so MGS can revoke it */
1582 if (!rcl)
1583 ldlm_lock_decref(&lockh, LCK_CR);
1584
1585 return rc;
1586 }
1587
1588
1589 /** Called from lustre_process_log.
1590 * LCFG_LOG_START gets the config log from the MGS, processes it to start
1591 * any services, and adds it to the list logs to watch (follow).
1592 */
1593 static int mgc_process_config(struct obd_device *obd, u32 len, void *buf)
1594 {
1595 struct lustre_cfg *lcfg = buf;
1596 struct config_llog_instance *cfg = NULL;
1597 char *logname;
1598 int rc = 0;
1599
1600 switch (lcfg->lcfg_command) {
1601 case LCFG_LOV_ADD_OBD: {
1602 /* Overloading this cfg command: register a new target */
1603 struct mgs_target_info *mti;
1604
1605 if (LUSTRE_CFG_BUFLEN(lcfg, 1) !=
1606 sizeof(struct mgs_target_info)) {
1607 rc = -EINVAL;
1608 goto out;
1609 }
1610
1611 mti = (struct mgs_target_info *)lustre_cfg_buf(lcfg, 1);
1612 CDEBUG(D_MGC, "add_target %s %#x\n",
1613 mti->mti_svname, mti->mti_flags);
1614 rc = mgc_target_register(obd->u.cli.cl_mgc_mgsexp, mti);
1615 break;
1616 }
1617 case LCFG_LOV_DEL_OBD:
1618 /* Unregister has no meaning at the moment. */
1619 CERROR("lov_del_obd unimplemented\n");
1620 rc = -ENOSYS;
1621 break;
1622 case LCFG_SPTLRPC_CONF: {
1623 rc = sptlrpc_process_config(lcfg);
1624 break;
1625 }
1626 case LCFG_LOG_START: {
1627 struct config_llog_data *cld;
1628 struct super_block *sb;
1629
1630 logname = lustre_cfg_string(lcfg, 1);
1631 cfg = (struct config_llog_instance *)lustre_cfg_buf(lcfg, 2);
1632 sb = *(struct super_block **)lustre_cfg_buf(lcfg, 3);
1633
1634 CDEBUG(D_MGC, "parse_log %s from %d\n", logname,
1635 cfg->cfg_last_idx);
1636
1637 /* We're only called through here on the initial mount */
1638 rc = config_log_add(obd, logname, cfg, sb);
1639 if (rc)
1640 break;
1641 cld = config_log_find(logname, cfg);
1642 if (cld == NULL) {
1643 rc = -ENOENT;
1644 break;
1645 }
1646
1647 /* COMPAT_146 */
1648 /* FIXME only set this for old logs! Right now this forces
1649 us to always skip the "inside markers" check */
1650 cld->cld_cfg.cfg_flags |= CFG_F_COMPAT146;
1651
1652 rc = mgc_process_log(obd, cld);
1653 if (rc == 0 && cld->cld_recover != NULL) {
1654 if (OCD_HAS_FLAG(&obd->u.cli.cl_import->
1655 imp_connect_data, IMP_RECOV)) {
1656 rc = mgc_process_log(obd, cld->cld_recover);
1657 } else {
1658 struct config_llog_data *cir = cld->cld_recover;
1659 cld->cld_recover = NULL;
1660 config_log_put(cir);
1661 }
1662 if (rc)
1663 CERROR("Cannot process recover llog %d\n", rc);
1664 }
1665
1666 if (rc == 0 && cld->cld_params != NULL) {
1667 rc = mgc_process_log(obd, cld->cld_params);
1668 if (rc == -ENOENT) {
1669 CDEBUG(D_MGC,
1670 "There is no params config file yet\n");
1671 rc = 0;
1672 }
1673 /* params log is optional */
1674 if (rc)
1675 CERROR(
1676 "%s: can't process params llog: rc = %d\n",
1677 obd->obd_name, rc);
1678 }
1679 config_log_put(cld);
1680
1681 break;
1682 }
1683 case LCFG_LOG_END: {
1684 logname = lustre_cfg_string(lcfg, 1);
1685
1686 if (lcfg->lcfg_bufcount >= 2)
1687 cfg = (struct config_llog_instance *)lustre_cfg_buf(
1688 lcfg, 2);
1689 rc = config_log_end(logname, cfg);
1690 break;
1691 }
1692 default: {
1693 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1694 rc = -EINVAL;
1695 goto out;
1696
1697 }
1698 }
1699 out:
1700 return rc;
1701 }
1702
1703 struct obd_ops mgc_obd_ops = {
1704 .o_owner = THIS_MODULE,
1705 .o_setup = mgc_setup,
1706 .o_precleanup = mgc_precleanup,
1707 .o_cleanup = mgc_cleanup,
1708 .o_add_conn = client_import_add_conn,
1709 .o_del_conn = client_import_del_conn,
1710 .o_connect = client_connect_import,
1711 .o_disconnect = client_disconnect_export,
1712 /* .o_enqueue = mgc_enqueue, */
1713 /* .o_iocontrol = mgc_iocontrol, */
1714 .o_set_info_async = mgc_set_info_async,
1715 .o_get_info = mgc_get_info,
1716 .o_import_event = mgc_import_event,
1717 .o_process_config = mgc_process_config,
1718 };
1719
1720 static int __init mgc_init(void)
1721 {
1722 return class_register_type(&mgc_obd_ops, NULL,
1723 LUSTRE_MGC_NAME, NULL);
1724 }
1725
1726 static void /*__exit*/ mgc_exit(void)
1727 {
1728 class_unregister_type(LUSTRE_MGC_NAME);
1729 }
1730
1731 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1732 MODULE_DESCRIPTION("Lustre Management Client");
1733 MODULE_LICENSE("GPL");
1734
1735 module_init(mgc_init);
1736 module_exit(mgc_exit);
Linux kernel - Deliverables
This page took 0.239827 seconds and 4 git commands to generate.