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, int count,
137 struct llog_cookie *cookies, int flags, void *data)
139 struct llog_canceld_ctxt *llcd;
145 if (ctxt->loc_imp == NULL) {
146 CWARN("no import for ctxt %p\n", ctxt);
150 if (count == 0 || cookies == NULL) {
151 down(&ctxt->loc_sem);
152 if (ctxt->loc_llcd == NULL || !(flags & OBD_LLOG_FL_SENDNOW))
155 llcd = ctxt->loc_llcd;
159 down(&ctxt->loc_sem);
160 llcd = ctxt->loc_llcd;
164 CERROR("couldn't get an llcd - dropped "LPX64":%x+%u\n",
165 cookies->lgc_lgl.lgl_oid,
166 cookies->lgc_lgl.lgl_ogen, cookies->lgc_index);
167 GOTO(out, rc = -ENOMEM);
169 llcd->llcd_ctxt = ctxt;
170 ctxt->loc_llcd = llcd;
173 memcpy((char *)llcd->llcd_cookies + llcd->llcd_cookiebytes, cookies,
175 llcd->llcd_cookiebytes += sizeof(*cookies);
178 if ((LLCD_SIZE - llcd->llcd_cookiebytes < sizeof(*cookies) ||
179 flags & OBD_LLOG_FL_SENDNOW)) {
180 CDEBUG(D_HA, "send llcd %p:%p\n", llcd, llcd->llcd_ctxt);
181 ctxt->loc_llcd = NULL;
188 EXPORT_SYMBOL(llog_obd_repl_cancel);
190 int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp)
195 if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) {
196 down(&ctxt->loc_sem);
197 if (ctxt->loc_llcd != NULL) {
198 CWARN("import will be destroyed, put "
199 "llcd %p:%p\n", ctxt->loc_llcd, ctxt);
200 llcd_put(ctxt->loc_llcd);
201 ctxt->loc_llcd = NULL;
202 ctxt->loc_imp = NULL;
206 rc = llog_cancel(ctxt, 0, NULL, OBD_LLOG_FL_SENDNOW, NULL);
211 EXPORT_SYMBOL(llog_obd_repl_sync);
213 static int log_commit_thread(void *arg)
215 struct llog_commit_master *lcm = arg;
216 struct llog_commit_daemon *lcd;
217 struct llog_canceld_ctxt *llcd, *n;
221 OBD_ALLOC(lcd, sizeof(*lcd));
226 ptlrpc_daemonize(); /* thread never needs to do IO */
228 SIGNAL_MASK_LOCK(current, flags);
229 sigfillset(¤t->blocked);
231 SIGNAL_MASK_UNLOCK(current, flags);
233 spin_lock(&lcm->lcm_thread_lock);
234 THREAD_NAME(current->comm, sizeof(current->comm) - 1,
235 "ll_log_comt_%02d", atomic_read(&lcm->lcm_thread_total));
236 atomic_inc(&lcm->lcm_thread_total);
237 spin_unlock(&lcm->lcm_thread_lock);
240 INIT_LIST_HEAD(&lcd->lcd_lcm_list);
241 INIT_LIST_HEAD(&lcd->lcd_llcd_list);
244 CDEBUG(D_HA, "%s started\n", current->comm);
246 struct ptlrpc_request *request;
247 struct obd_import *import = NULL;
248 struct list_head *sending_list;
251 /* If we do not have enough pages available, allocate some */
252 while (atomic_read(&lcm->lcm_llcd_numfree) <
253 lcm->lcm_llcd_minfree) {
254 if (llcd_alloc() < 0)
258 spin_lock(&lcm->lcm_thread_lock);
259 atomic_inc(&lcm->lcm_thread_numidle);
260 list_move(&lcd->lcd_lcm_list, &lcm->lcm_thread_idle);
261 spin_unlock(&lcm->lcm_thread_lock);
263 wait_event_interruptible(lcm->lcm_waitq,
264 !list_empty(&lcm->lcm_llcd_pending) ||
265 lcm->lcm_flags & LLOG_LCM_FL_EXIT);
267 /* If we are the last available thread, start a new one in case
268 * we get blocked on an RPC (nobody else will start a new one)*/
269 spin_lock(&lcm->lcm_thread_lock);
270 atomic_dec(&lcm->lcm_thread_numidle);
271 list_move(&lcd->lcd_lcm_list, &lcm->lcm_thread_busy);
272 spin_unlock(&lcm->lcm_thread_lock);
274 sending_list = &lcm->lcm_llcd_pending;
276 if (lcm->lcm_flags & LLOG_LCM_FL_EXIT) {
277 lcm->lcm_llcd_maxfree = 0;
278 lcm->lcm_llcd_minfree = 0;
279 lcm->lcm_thread_max = 0;
281 if (list_empty(&lcm->lcm_llcd_pending) ||
282 lcm->lcm_flags & LLOG_LCM_FL_EXIT_FORCE)
286 if (atomic_read(&lcm->lcm_thread_numidle) <= 1 &&
287 atomic_read(&lcm->lcm_thread_total) < lcm->lcm_thread_max) {
288 rc = llog_start_commit_thread();
290 CERROR("error starting thread: rc %d\n", rc);
293 /* Move all of the pending cancels from the same OST off of
294 * the list, so we don't get multiple threads blocked and/or
295 * doing upcalls on the same OST in case of failure. */
296 spin_lock(&lcm->lcm_llcd_lock);
297 if (!list_empty(sending_list)) {
298 list_move_tail(sending_list->next,
299 &lcd->lcd_llcd_list);
300 llcd = list_entry(lcd->lcd_llcd_list.next,
301 typeof(*llcd), llcd_list);
302 LASSERT(llcd->llcd_lcm == lcm);
303 import = llcd->llcd_ctxt->loc_imp;
305 list_for_each_entry_safe(llcd, n, sending_list, llcd_list) {
306 LASSERT(llcd->llcd_lcm == lcm);
307 if (import == llcd->llcd_ctxt->loc_imp)
308 list_move_tail(&llcd->llcd_list,
309 &lcd->lcd_llcd_list);
311 if (sending_list != &lcm->lcm_llcd_resend) {
312 list_for_each_entry_safe(llcd, n, &lcm->lcm_llcd_resend,
314 LASSERT(llcd->llcd_lcm == lcm);
315 if (import == llcd->llcd_ctxt->loc_imp)
316 list_move_tail(&llcd->llcd_list,
317 &lcd->lcd_llcd_list);
320 spin_unlock(&lcm->lcm_llcd_lock);
322 /* We are the only one manipulating our local list - no lock */
323 list_for_each_entry_safe(llcd,n, &lcd->lcd_llcd_list,llcd_list){
324 char *bufs[1] = {(char *)llcd->llcd_cookies};
326 list_del(&llcd->llcd_list);
327 if (llcd->llcd_cookiebytes == 0) {
328 CDEBUG(D_HA, "put empty llcd %p:%p\n",
329 llcd, llcd->llcd_ctxt);
334 down(&llcd->llcd_ctxt->loc_sem);
335 if (llcd->llcd_ctxt->loc_imp == NULL) {
336 up(&llcd->llcd_ctxt->loc_sem);
337 CWARN("import will be destroyed, put "
338 "llcd %p:%p\n", llcd, llcd->llcd_ctxt);
343 request = ptlrpc_prep_req(import, OBD_LOG_CANCEL, 1,
344 &llcd->llcd_cookiebytes,
346 up(&llcd->llcd_ctxt->loc_sem);
348 if (request == NULL) {
350 CERROR("error preparing commit: rc %d\n", rc);
352 spin_lock(&lcm->lcm_llcd_lock);
353 list_splice(&lcd->lcd_llcd_list,
354 &lcm->lcm_llcd_resend);
355 INIT_LIST_HEAD(&lcd->lcd_llcd_list);
356 spin_unlock(&lcm->lcm_llcd_lock);
360 request->rq_replen = lustre_msg_size(0, NULL);
361 down(&llcd->llcd_ctxt->loc_sem);
362 if (llcd->llcd_ctxt->loc_imp == NULL) {
363 up(&llcd->llcd_ctxt->loc_sem);
364 CWARN("import will be destroyed, put "
365 "llcd %p:%p\n", llcd, llcd->llcd_ctxt);
367 ptlrpc_req_finished(request);
370 rc = ptlrpc_queue_wait(request);
371 ptlrpc_req_finished(request);
372 up(&llcd->llcd_ctxt->loc_sem);
374 /* If the RPC failed, we put this and the remaining
375 * messages onto the resend list for another time. */
381 #if 0 /* FIXME just put llcd, not put it on resend list */
382 spin_lock(&lcm->lcm_llcd_lock);
383 list_splice(&lcd->lcd_llcd_list, &lcm->lcm_llcd_resend);
384 if (++llcd->llcd_tries < 5) {
385 CERROR("commit %p failed on attempt %d: rc %d\n",
386 llcd, llcd->llcd_tries, rc);
388 list_add_tail(&llcd->llcd_list,
389 &lcm->lcm_llcd_resend);
390 spin_unlock(&lcm->lcm_llcd_lock);
392 spin_unlock(&lcm->lcm_llcd_lock);
394 CERROR("commit %p:%p drop %d cookies: rc %d\n",
395 llcd, llcd->llcd_ctxt,
396 (int)(llcd->llcd_cookiebytes /
397 sizeof(*llcd->llcd_cookies)), rc);
406 sending_list = &lcm->lcm_llcd_resend;
407 if (!list_empty(sending_list))
412 /* If we are force exiting, just drop all of the cookies. */
413 if (lcm->lcm_flags & LLOG_LCM_FL_EXIT_FORCE) {
414 spin_lock(&lcm->lcm_llcd_lock);
415 list_splice(&lcm->lcm_llcd_pending, &lcd->lcd_llcd_list);
416 list_splice(&lcm->lcm_llcd_resend, &lcd->lcd_llcd_list);
417 list_splice(&lcm->lcm_llcd_free, &lcd->lcd_llcd_list);
418 spin_unlock(&lcm->lcm_llcd_lock);
420 list_for_each_entry_safe(llcd, n, &lcd->lcd_llcd_list,llcd_list)
424 spin_lock(&lcm->lcm_thread_lock);
425 list_del(&lcd->lcd_lcm_list);
426 spin_unlock(&lcm->lcm_thread_lock);
427 OBD_FREE(lcd, sizeof(*lcd));
429 spin_lock(&lcm->lcm_thread_lock);
430 atomic_dec(&lcm->lcm_thread_total);
431 spin_unlock(&lcm->lcm_thread_lock);
432 wake_up(&lcm->lcm_waitq);
434 CDEBUG(D_HA, "%s exiting\n", current->comm);
438 int llog_start_commit_thread(void)
443 if (atomic_read(&lcm->lcm_thread_total) >= lcm->lcm_thread_max)
446 rc = kernel_thread(log_commit_thread, lcm, CLONE_VM | CLONE_FILES);
448 CERROR("error starting thread #%d: %d\n",
449 atomic_read(&lcm->lcm_thread_total), rc);
455 EXPORT_SYMBOL(llog_start_commit_thread);
457 static struct llog_process_args {
458 struct semaphore llpa_sem;
459 struct llog_ctxt *llpa_ctxt;
464 int llog_init_commit_master(void)
466 INIT_LIST_HEAD(&lcm->lcm_thread_busy);
467 INIT_LIST_HEAD(&lcm->lcm_thread_idle);
468 spin_lock_init(&lcm->lcm_thread_lock);
469 atomic_set(&lcm->lcm_thread_numidle, 0);
470 init_waitqueue_head(&lcm->lcm_waitq);
471 INIT_LIST_HEAD(&lcm->lcm_llcd_pending);
472 INIT_LIST_HEAD(&lcm->lcm_llcd_resend);
473 INIT_LIST_HEAD(&lcm->lcm_llcd_free);
474 spin_lock_init(&lcm->lcm_llcd_lock);
475 atomic_set(&lcm->lcm_llcd_numfree, 0);
476 lcm->lcm_llcd_minfree = 0;
477 lcm->lcm_thread_max = 5;
478 /* FIXME initialize semaphore for llog_process_args */
479 sema_init(&llpa.llpa_sem, 1);
483 int llog_cleanup_commit_master(int force)
485 lcm->lcm_flags |= LLOG_LCM_FL_EXIT;
487 lcm->lcm_flags |= LLOG_LCM_FL_EXIT_FORCE;
488 wake_up(&lcm->lcm_waitq);
490 wait_event_interruptible(lcm->lcm_waitq,
491 atomic_read(&lcm->lcm_thread_total) == 0);
495 static int log_process_thread(void *args)
497 struct llog_process_args *data = args;
498 struct llog_ctxt *ctxt = data->llpa_ctxt;
499 void *cb = data->llpa_cb;
500 struct llog_logid logid = *(struct llog_logid *)(data->llpa_arg);
501 struct llog_handle *llh = NULL;
508 ptlrpc_daemonize(); /* thread never needs to do IO */
510 SIGNAL_MASK_LOCK(current, flags);
511 sigfillset(¤t->blocked);
513 SIGNAL_MASK_UNLOCK(current, flags);
516 rc = llog_create(ctxt, &llh, &logid, NULL);
518 CERROR("llog_create failed %d\n", rc);
521 rc = llog_init_handle(llh, LLOG_F_IS_CAT, NULL);
523 CERROR("llog_init_handle failed %d\n", rc);
528 rc = llog_cat_process(llh, (llog_cb_t)cb, NULL);
529 if (rc != LLOG_PROC_BREAK)
530 CERROR("llog_cat_process failed %d\n", rc);
532 CWARN("no callback function for recovery\n");
535 CDEBUG(D_HA, "send llcd %p:%p forcibly after recovery\n",
536 ctxt->loc_llcd, ctxt);
537 llog_sync(ctxt, NULL);
539 rc = llog_cat_put(llh);
541 CERROR("llog_cat_put failed %d\n", rc);
546 static int llog_recovery_generic(struct llog_ctxt *ctxt, void *handle,void *arg)
551 down(&llpa.llpa_sem);
552 llpa.llpa_ctxt = ctxt;
553 llpa.llpa_cb = handle;
556 rc = kernel_thread(log_process_thread, &llpa, CLONE_VM | CLONE_FILES);
558 CERROR("error starting log_process_thread: %d\n", rc);
560 CDEBUG(D_HA, "log_process_thread: %d\n", rc);
567 int llog_repl_connect(struct llog_ctxt *ctxt, int count,
568 struct llog_logid *logid, struct llog_gen *gen,
569 struct obd_uuid *uuid)
571 struct llog_canceld_ctxt *llcd;
575 /* send back llcd before recovery from llog */
576 if (ctxt->loc_llcd != NULL) {
577 CWARN("llcd %p:%p not empty\n", ctxt->loc_llcd, ctxt);
578 llog_sync(ctxt, NULL);
581 down(&ctxt->loc_sem);
582 ctxt->loc_gen = *gen;
585 CERROR("couldn't get an llcd\n");
589 llcd->llcd_ctxt = ctxt;
590 ctxt->loc_llcd = llcd;
593 rc = llog_recovery_generic(ctxt, ctxt->loc_proc_cb, logid);
595 CERROR("error recovery process: %d\n", rc);
599 EXPORT_SYMBOL(llog_repl_connect);
601 #else /* !__KERNEL__ */
603 int llog_obd_repl_cancel(struct llog_ctxt *ctxt, int count,
604 struct llog_cookie *cookies, int flags, void *data)