1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2003 Cluster File Systems, Inc.
5 * Author: Andreas Dilger <adilger@clusterfs.com>
7 * This file is part of Lustre, http://www.lustre.org.
9 * Lustre is free software; you can redistribute it and/or
10 * modify it under the terms of version 2 of the GNU General Public
11 * License as published by the Free Software Foundation.
13 * Lustre is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with Lustre; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
22 * OST<->MDS recovery logging thread.
24 * Invariants in implementation:
25 * - we do not share logs among different OST<->MDS connections, so that
26 * if an OST or MDS fails it need only look at log(s) relevant to itself
29 #define DEBUG_SUBSYSTEM S_LOG
32 # define EXPORT_SYMTAB
38 # include <portals/list.h>
39 # include <liblustre.h>
42 #include <linux/kp30.h>
43 #include <linux/obd_class.h>
44 #include <linux/lustre_commit_confd.h>
45 #include <linux/obd_support.h>
46 #include <linux/obd_class.h>
47 #include <linux/lustre_net.h>
48 #include <portals/types.h>
49 #include <portals/list.h>
50 #include <linux/lustre_log.h>
51 #include "ptlrpc_internal.h"
53 #define LLCD_SIZE 4096
57 static struct llog_commit_master lustre_lcm;
58 static struct llog_commit_master *lcm = &lustre_lcm;
60 /* Allocate new commit structs in case we do not have enough */
61 static int llcd_alloc(void)
63 struct llog_canceld_ctxt *llcd;
64 int offset = offsetof(struct llog_canceld_ctxt, llcd_cookies);
66 OBD_ALLOC(llcd, LLCD_SIZE + offset);
72 spin_lock(&lcm->lcm_llcd_lock);
73 list_add(&llcd->llcd_list, &lcm->lcm_llcd_free);
74 atomic_inc(&lcm->lcm_llcd_numfree);
75 spin_unlock(&lcm->lcm_llcd_lock);
80 /* Get a free cookie struct from the list */
81 struct llog_canceld_ctxt *llcd_grab(void)
83 struct llog_canceld_ctxt *llcd;
85 spin_lock(&lcm->lcm_llcd_lock);
86 if (list_empty(&lcm->lcm_llcd_free)) {
87 spin_unlock(&lcm->lcm_llcd_lock);
88 if (llcd_alloc() < 0) {
89 CERROR("unable to allocate log commit data!\n");
92 spin_lock(&lcm->lcm_llcd_lock);
95 llcd = list_entry(lcm->lcm_llcd_free.next, typeof(*llcd), llcd_list);
96 list_del(&llcd->llcd_list);
97 atomic_dec(&lcm->lcm_llcd_numfree);
98 spin_unlock(&lcm->lcm_llcd_lock);
100 llcd->llcd_tries = 0;
101 llcd->llcd_cookiebytes = 0;
105 EXPORT_SYMBOL(llcd_grab);
107 static void llcd_put(struct llog_canceld_ctxt *llcd)
109 int offset = offsetof(struct llog_canceld_ctxt, llcd_cookies);
111 if (atomic_read(&lcm->lcm_llcd_numfree) >= lcm->lcm_llcd_maxfree) {
112 OBD_FREE(llcd, LLCD_SIZE + offset);
114 spin_lock(&lcm->lcm_llcd_lock);
115 list_add(&llcd->llcd_list, &lcm->lcm_llcd_free);
116 atomic_inc(&lcm->lcm_llcd_numfree);
117 spin_unlock(&lcm->lcm_llcd_lock);
121 /* Send some cookies to the appropriate target */
122 void llcd_send(struct llog_canceld_ctxt *llcd)
124 spin_lock(&llcd->llcd_lcm->lcm_llcd_lock);
125 list_add_tail(&llcd->llcd_list, &llcd->llcd_lcm->lcm_llcd_pending);
126 spin_unlock(&llcd->llcd_lcm->lcm_llcd_lock);
128 wake_up_nr(&llcd->llcd_lcm->lcm_waitq, 1);
130 EXPORT_SYMBOL(llcd_send);
132 /* deleted objects have a commit callback that cancels the MDS
133 * log record for the deletion. The commit callback calls this
136 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
137 struct lov_stripe_md *lsm, int count,
138 struct llog_cookie *cookies, int flags)
140 struct llog_canceld_ctxt *llcd;
146 if (count == 0 || cookies == NULL) {
147 down(&ctxt->loc_sem);
148 if (ctxt->loc_llcd == NULL || !(flags & OBD_LLOG_FL_SENDNOW))
151 llcd = ctxt->loc_llcd;
155 down(&ctxt->loc_sem);
156 llcd = ctxt->loc_llcd;
160 CERROR("couldn't get an llcd - dropped "LPX64":%x+%u\n",
161 cookies->lgc_lgl.lgl_oid,
162 cookies->lgc_lgl.lgl_ogen, cookies->lgc_index);
163 GOTO(out, rc = -ENOMEM);
165 llcd->llcd_ctxt = ctxt;
166 ctxt->loc_llcd = llcd;
169 memcpy((char *)llcd->llcd_cookies + llcd->llcd_cookiebytes, cookies,
171 llcd->llcd_cookiebytes += sizeof(*cookies);
174 if ((LLCD_SIZE - llcd->llcd_cookiebytes < sizeof(*cookies) ||
175 flags & OBD_LLOG_FL_SENDNOW)) {
176 CDEBUG(D_HA, "send llcd %p:%p\n", llcd, llcd->llcd_ctxt);
177 ctxt->loc_llcd = NULL;
184 EXPORT_SYMBOL(llog_obd_repl_cancel);
186 int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp)
191 if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) {
192 down(&ctxt->loc_sem);
193 if (ctxt->loc_llcd != NULL) {
194 CWARN("import will be destroyed, put "
195 "llcd %p:%p\n", ctxt->loc_llcd, ctxt);
196 llcd_put(ctxt->loc_llcd);
197 ctxt->loc_llcd = NULL;
198 ctxt->loc_imp = NULL;
202 rc = llog_cancel(ctxt, NULL, 0, NULL, OBD_LLOG_FL_SENDNOW);
207 EXPORT_SYMBOL(llog_obd_repl_sync);
209 static int log_commit_thread(void *arg)
211 struct llog_commit_master *lcm = arg;
212 struct llog_commit_daemon *lcd;
213 struct llog_canceld_ctxt *llcd, *n;
217 OBD_ALLOC(lcd, sizeof(*lcd));
222 ptlrpc_daemonize(); /* thread never needs to do IO */
224 SIGNAL_MASK_LOCK(current, flags);
225 sigfillset(¤t->blocked);
227 SIGNAL_MASK_UNLOCK(current, flags);
229 spin_lock(&lcm->lcm_thread_lock);
230 THREAD_NAME(current->comm, "ll_log_commit_%d",
231 atomic_read(&lcm->lcm_thread_total));
232 atomic_inc(&lcm->lcm_thread_total);
233 spin_unlock(&lcm->lcm_thread_lock);
236 INIT_LIST_HEAD(&lcd->lcd_lcm_list);
237 INIT_LIST_HEAD(&lcd->lcd_llcd_list);
240 CDEBUG(D_HA, "%s started\n", current->comm);
242 struct ptlrpc_request *request;
243 struct obd_import *import = NULL;
244 struct list_head *sending_list;
247 /* If we do not have enough pages available, allocate some */
248 while (atomic_read(&lcm->lcm_llcd_numfree) <
249 lcm->lcm_llcd_minfree) {
250 if (llcd_alloc() < 0)
254 spin_lock(&lcm->lcm_thread_lock);
255 atomic_inc(&lcm->lcm_thread_numidle);
256 list_move(&lcd->lcd_lcm_list, &lcm->lcm_thread_idle);
257 spin_unlock(&lcm->lcm_thread_lock);
259 wait_event_interruptible(lcm->lcm_waitq,
260 !list_empty(&lcm->lcm_llcd_pending) ||
261 lcm->lcm_flags & LLOG_LCM_FL_EXIT);
263 /* If we are the last available thread, start a new one in case
264 * we get blocked on an RPC (nobody else will start a new one)*/
265 spin_lock(&lcm->lcm_thread_lock);
266 atomic_dec(&lcm->lcm_thread_numidle);
267 list_move(&lcd->lcd_lcm_list, &lcm->lcm_thread_busy);
268 spin_unlock(&lcm->lcm_thread_lock);
270 sending_list = &lcm->lcm_llcd_pending;
272 if (lcm->lcm_flags & LLOG_LCM_FL_EXIT) {
273 lcm->lcm_llcd_maxfree = 0;
274 lcm->lcm_llcd_minfree = 0;
275 lcm->lcm_thread_max = 0;
277 if (list_empty(&lcm->lcm_llcd_pending) ||
278 lcm->lcm_flags & LLOG_LCM_FL_EXIT_FORCE)
282 if (atomic_read(&lcm->lcm_thread_numidle) <= 1 &&
283 atomic_read(&lcm->lcm_thread_total) < lcm->lcm_thread_max) {
284 rc = llog_start_commit_thread();
286 CERROR("error starting thread: rc %d\n", rc);
289 /* Move all of the pending cancels from the same OST off of
290 * the list, so we don't get multiple threads blocked and/or
291 * doing upcalls on the same OST in case of failure. */
292 spin_lock(&lcm->lcm_llcd_lock);
293 if (!list_empty(sending_list)) {
294 list_move_tail(sending_list->next,
295 &lcd->lcd_llcd_list);
296 llcd = list_entry(lcd->lcd_llcd_list.next,
297 typeof(*llcd), llcd_list);
298 LASSERT(llcd->llcd_lcm == lcm);
299 import = llcd->llcd_ctxt->loc_imp;
301 list_for_each_entry_safe(llcd, n, sending_list, llcd_list) {
302 LASSERT(llcd->llcd_lcm == lcm);
303 if (import == llcd->llcd_ctxt->loc_imp)
304 list_move_tail(&llcd->llcd_list,
305 &lcd->lcd_llcd_list);
307 if (sending_list != &lcm->lcm_llcd_resend) {
308 list_for_each_entry_safe(llcd, n, &lcm->lcm_llcd_resend,
310 LASSERT(llcd->llcd_lcm == lcm);
311 if (import == llcd->llcd_ctxt->loc_imp)
312 list_move_tail(&llcd->llcd_list,
313 &lcd->lcd_llcd_list);
316 spin_unlock(&lcm->lcm_llcd_lock);
318 /* We are the only one manipulating our local list - no lock */
319 list_for_each_entry_safe(llcd,n, &lcd->lcd_llcd_list,llcd_list){
320 char *bufs[1] = {(char *)llcd->llcd_cookies};
322 list_del(&llcd->llcd_list);
323 if (llcd->llcd_cookiebytes == 0) {
324 CDEBUG(D_HA, "put empty llcd %p:%p\n",
325 llcd, llcd->llcd_ctxt);
330 down(&llcd->llcd_ctxt->loc_sem);
331 if (llcd->llcd_ctxt->loc_imp == NULL) {
332 up(&llcd->llcd_ctxt->loc_sem);
333 CWARN("import will be destroyed, put "
334 "llcd %p:%p\n", llcd, llcd->llcd_ctxt);
339 request = ptlrpc_prep_req(import, OBD_LOG_CANCEL, 1,
340 &llcd->llcd_cookiebytes,
342 up(&llcd->llcd_ctxt->loc_sem);
344 if (request == NULL) {
346 CERROR("error preparing commit: rc %d\n", rc);
348 spin_lock(&lcm->lcm_llcd_lock);
349 list_splice(&lcd->lcd_llcd_list,
350 &lcm->lcm_llcd_resend);
351 INIT_LIST_HEAD(&lcd->lcd_llcd_list);
352 spin_unlock(&lcm->lcm_llcd_lock);
356 request->rq_replen = lustre_msg_size(0, NULL);
357 down(&llcd->llcd_ctxt->loc_sem);
358 if (llcd->llcd_ctxt->loc_imp == NULL) {
359 up(&llcd->llcd_ctxt->loc_sem);
360 CWARN("import will be destroyed, put "
361 "llcd %p:%p\n", llcd, llcd->llcd_ctxt);
363 ptlrpc_req_finished(request);
366 rc = ptlrpc_queue_wait(request);
367 ptlrpc_req_finished(request);
368 up(&llcd->llcd_ctxt->loc_sem);
370 /* If the RPC failed, we put this and the remaining
371 * messages onto the resend list for another time. */
377 #if 0 /* FIXME just put llcd, not put it on resend list */
378 spin_lock(&lcm->lcm_llcd_lock);
379 list_splice(&lcd->lcd_llcd_list, &lcm->lcm_llcd_resend);
380 if (++llcd->llcd_tries < 5) {
381 CERROR("commit %p failed on attempt %d: rc %d\n",
382 llcd, llcd->llcd_tries, rc);
384 list_add_tail(&llcd->llcd_list,
385 &lcm->lcm_llcd_resend);
386 spin_unlock(&lcm->lcm_llcd_lock);
388 spin_unlock(&lcm->lcm_llcd_lock);
390 CERROR("commit %p:%p drop %d cookies: rc %d\n",
391 llcd, llcd->llcd_ctxt,
392 (int)(llcd->llcd_cookiebytes /
393 sizeof(*llcd->llcd_cookies)), rc);
402 sending_list = &lcm->lcm_llcd_resend;
403 if (!list_empty(sending_list))
408 /* If we are force exiting, just drop all of the cookies. */
409 if (lcm->lcm_flags & LLOG_LCM_FL_EXIT_FORCE) {
410 spin_lock(&lcm->lcm_llcd_lock);
411 list_splice(&lcm->lcm_llcd_pending, &lcd->lcd_llcd_list);
412 list_splice(&lcm->lcm_llcd_resend, &lcd->lcd_llcd_list);
413 list_splice(&lcm->lcm_llcd_free, &lcd->lcd_llcd_list);
414 spin_unlock(&lcm->lcm_llcd_lock);
416 list_for_each_entry_safe(llcd, n, &lcd->lcd_llcd_list,llcd_list)
420 spin_lock(&lcm->lcm_thread_lock);
421 list_del(&lcd->lcd_lcm_list);
422 spin_unlock(&lcm->lcm_thread_lock);
423 OBD_FREE(lcd, sizeof(*lcd));
425 spin_lock(&lcm->lcm_thread_lock);
426 atomic_dec(&lcm->lcm_thread_total);
427 spin_unlock(&lcm->lcm_thread_lock);
428 wake_up(&lcm->lcm_waitq);
430 CDEBUG(D_HA, "%s exiting\n", current->comm);
434 int llog_start_commit_thread(void)
439 if (atomic_read(&lcm->lcm_thread_total) >= lcm->lcm_thread_max)
442 rc = kernel_thread(log_commit_thread, lcm, CLONE_VM | CLONE_FILES);
444 CERROR("error starting thread #%d: %d\n",
445 atomic_read(&lcm->lcm_thread_total), rc);
451 EXPORT_SYMBOL(llog_start_commit_thread);
453 static struct llog_process_args {
454 struct semaphore llpa_sem;
455 struct llog_ctxt *llpa_ctxt;
460 int llog_init_commit_master(void)
462 INIT_LIST_HEAD(&lcm->lcm_thread_busy);
463 INIT_LIST_HEAD(&lcm->lcm_thread_idle);
464 spin_lock_init(&lcm->lcm_thread_lock);
465 atomic_set(&lcm->lcm_thread_numidle, 0);
466 init_waitqueue_head(&lcm->lcm_waitq);
467 INIT_LIST_HEAD(&lcm->lcm_llcd_pending);
468 INIT_LIST_HEAD(&lcm->lcm_llcd_resend);
469 INIT_LIST_HEAD(&lcm->lcm_llcd_free);
470 spin_lock_init(&lcm->lcm_llcd_lock);
471 atomic_set(&lcm->lcm_llcd_numfree, 0);
472 lcm->lcm_llcd_minfree = 0;
473 lcm->lcm_thread_max = 5;
474 /* FIXME initialize semaphore for llog_process_args */
475 sema_init(&llpa.llpa_sem, 1);
479 int llog_cleanup_commit_master(int force)
481 lcm->lcm_flags |= LLOG_LCM_FL_EXIT;
483 lcm->lcm_flags |= LLOG_LCM_FL_EXIT_FORCE;
484 wake_up(&lcm->lcm_waitq);
486 wait_event_interruptible(lcm->lcm_waitq,
487 atomic_read(&lcm->lcm_thread_total) == 0);
491 static int log_process_thread(void *args)
493 struct llog_process_args *data = args;
494 struct llog_ctxt *ctxt = data->llpa_ctxt;
495 void *cb = data->llpa_cb;
496 struct llog_logid logid = *(struct llog_logid *)(data->llpa_arg);
497 struct llog_handle *llh = NULL;
504 ptlrpc_daemonize(); /* thread never needs to do IO */
506 SIGNAL_MASK_LOCK(current, flags);
507 sigfillset(¤t->blocked);
509 SIGNAL_MASK_UNLOCK(current, flags);
512 rc = llog_create(ctxt, &llh, &logid, NULL);
514 CERROR("llog_create failed %d\n", rc);
517 rc = llog_init_handle(llh, LLOG_F_IS_CAT, NULL);
519 CERROR("llog_init_handle failed %d\n", rc);
524 rc = llog_cat_process(llh, (llog_cb_t)cb, NULL);
525 if (rc != LLOG_PROC_BREAK)
526 CERROR("llog_cat_process failed %d\n", rc);
528 CWARN("no callback function for recovery\n");
531 CDEBUG(D_HA, "send llcd %p:%p forcibly after recovery\n",
532 ctxt->loc_llcd, ctxt);
533 llog_sync(ctxt, NULL);
535 rc = llog_cat_put(llh);
537 CERROR("llog_cat_put failed %d\n", rc);
542 static int llog_recovery_generic(struct llog_ctxt *ctxt, void *handle,void *arg)
547 down(&llpa.llpa_sem);
548 llpa.llpa_ctxt = ctxt;
549 llpa.llpa_cb = handle;
552 rc = kernel_thread(log_process_thread, &llpa, CLONE_VM | CLONE_FILES);
554 CERROR("error starting log_process_thread: %d\n", rc);
556 CDEBUG(D_HA, "log_process_thread: %d\n", rc);
563 int llog_repl_connect(struct llog_ctxt *ctxt, int count,
564 struct llog_logid *logid, struct llog_gen *gen)
566 struct llog_canceld_ctxt *llcd;
570 /* send back llcd before recovery from llog */
571 if (ctxt->loc_llcd != NULL) {
572 CWARN("llcd %p:%p not empty\n", ctxt->loc_llcd, ctxt);
573 llog_sync(ctxt, NULL);
576 down(&ctxt->loc_sem);
577 ctxt->loc_gen = *gen;
580 CERROR("couldn't get an llcd\n");
584 llcd->llcd_ctxt = ctxt;
585 ctxt->loc_llcd = llcd;
588 rc = llog_recovery_generic(ctxt, ctxt->llog_proc_cb, logid);
590 CERROR("error recovery process: %d\n", rc);
594 EXPORT_SYMBOL(llog_repl_connect);
596 #else /* !__KERNEL__ */
598 int llog_obd_repl_cancel(struct llog_ctxt *ctxt,
599 struct lov_stripe_md *lsm, int count,
600 struct llog_cookie *cookies, int flags)