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 <libcfs/list.h>
39 # include <liblustre.h>
42 #include <libcfs/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 <libcfs/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;
144 down(&ctxt->loc_sem);
145 if (ctxt->loc_imp == NULL) {
146 CWARN("no import for ctxt %p\n", ctxt);
150 llcd = ctxt->loc_llcd;
151 if (count > 0 && cookies != NULL) {
155 CERROR("couldn't get an llcd - dropped "LPX64":%x+%u\n",
156 cookies->lgc_lgl.lgl_oid,
157 cookies->lgc_lgl.lgl_ogen, cookies->lgc_index);
158 GOTO(out, rc = -ENOMEM);
160 llcd->llcd_ctxt = ctxt;
161 ctxt->loc_llcd = llcd;
163 memcpy((char *)llcd->llcd_cookies + llcd->llcd_cookiebytes,
164 cookies, sizeof(*cookies));
165 llcd->llcd_cookiebytes += sizeof(*cookies);
167 if (llcd == NULL || !(flags & OBD_LLOG_FL_SENDNOW))
171 if ((LLCD_SIZE - llcd->llcd_cookiebytes < sizeof(*cookies) ||
172 flags & OBD_LLOG_FL_SENDNOW)) {
173 CDEBUG(D_HA, "send llcd %p:%p\n", llcd, llcd->llcd_ctxt);
174 ctxt->loc_llcd = NULL;
181 EXPORT_SYMBOL(llog_obd_repl_cancel);
183 int llog_obd_repl_sync(struct llog_ctxt *ctxt, struct obd_export *exp)
188 if (exp && (ctxt->loc_imp == exp->exp_imp_reverse)) {
189 down(&ctxt->loc_sem);
190 if (ctxt->loc_llcd != NULL) {
191 llcd_put(ctxt->loc_llcd);
192 ctxt->loc_llcd = NULL;
194 CWARN("reverse import disconnected, put "
195 "llcd %p:%p\n", ctxt->loc_llcd, ctxt);
196 ctxt->loc_imp = NULL;
199 rc = llog_cancel(ctxt, 0, NULL, OBD_LLOG_FL_SENDNOW, NULL);
204 EXPORT_SYMBOL(llog_obd_repl_sync);
206 static int log_commit_thread(void *arg)
208 struct llog_commit_master *lcm = arg;
209 struct llog_commit_daemon *lcd;
210 struct llog_canceld_ctxt *llcd, *n;
214 OBD_ALLOC(lcd, sizeof(*lcd));
219 ptlrpc_daemonize(); /* thread never needs to do IO */
221 SIGNAL_MASK_LOCK(current, flags);
222 sigfillset(¤t->blocked);
224 SIGNAL_MASK_UNLOCK(current, flags);
226 spin_lock(&lcm->lcm_thread_lock);
228 THREAD_NAME(current->comm, sizeof(current->comm) - 1,
229 "ll_log_comt_%02d", atomic_read(&lcm->lcm_thread_total));
230 atomic_inc(&lcm->lcm_thread_total);
231 spin_unlock(&lcm->lcm_thread_lock);
234 INIT_LIST_HEAD(&lcd->lcd_lcm_list);
235 INIT_LIST_HEAD(&lcd->lcd_llcd_list);
238 CDEBUG(D_HA, "%s started\n", current->comm);
240 struct ptlrpc_request *request;
241 struct obd_import *import = NULL;
242 struct list_head *sending_list;
245 /* If we do not have enough pages available, allocate some */
246 while (atomic_read(&lcm->lcm_llcd_numfree) <
247 lcm->lcm_llcd_minfree) {
248 if (llcd_alloc() < 0)
252 spin_lock(&lcm->lcm_thread_lock);
253 atomic_inc(&lcm->lcm_thread_numidle);
254 list_move(&lcd->lcd_lcm_list, &lcm->lcm_thread_idle);
255 spin_unlock(&lcm->lcm_thread_lock);
257 wait_event_interruptible(lcm->lcm_waitq,
258 !list_empty(&lcm->lcm_llcd_pending) ||
259 lcm->lcm_flags & LLOG_LCM_FL_EXIT);
261 /* If we are the last available thread, start a new one in case
262 * we get blocked on an RPC (nobody else will start a new one)*/
263 spin_lock(&lcm->lcm_thread_lock);
264 atomic_dec(&lcm->lcm_thread_numidle);
265 list_move(&lcd->lcd_lcm_list, &lcm->lcm_thread_busy);
266 spin_unlock(&lcm->lcm_thread_lock);
268 sending_list = &lcm->lcm_llcd_pending;
270 if (lcm->lcm_flags & LLOG_LCM_FL_EXIT) {
271 lcm->lcm_llcd_maxfree = 0;
272 lcm->lcm_llcd_minfree = 0;
273 lcm->lcm_thread_max = 0;
275 if (list_empty(&lcm->lcm_llcd_pending) ||
276 lcm->lcm_flags & LLOG_LCM_FL_EXIT_FORCE)
280 if (atomic_read(&lcm->lcm_thread_numidle) <= 1 &&
281 atomic_read(&lcm->lcm_thread_total) < lcm->lcm_thread_max) {
282 rc = llog_start_commit_thread();
284 CERROR("error starting thread: rc %d\n", rc);
287 /* Move all of the pending cancels from the same OST off of
288 * the list, so we don't get multiple threads blocked and/or
289 * doing upcalls on the same OST in case of failure. */
290 spin_lock(&lcm->lcm_llcd_lock);
291 if (!list_empty(sending_list)) {
292 list_move_tail(sending_list->next,
293 &lcd->lcd_llcd_list);
294 llcd = list_entry(lcd->lcd_llcd_list.next,
295 typeof(*llcd), llcd_list);
296 LASSERT(llcd->llcd_lcm == lcm);
297 import = llcd->llcd_ctxt->loc_imp;
299 list_for_each_entry_safe(llcd, n, sending_list, llcd_list) {
300 LASSERT(llcd->llcd_lcm == lcm);
301 if (import == llcd->llcd_ctxt->loc_imp)
302 list_move_tail(&llcd->llcd_list,
303 &lcd->lcd_llcd_list);
305 if (sending_list != &lcm->lcm_llcd_resend) {
306 list_for_each_entry_safe(llcd, n, &lcm->lcm_llcd_resend,
308 LASSERT(llcd->llcd_lcm == lcm);
309 if (import == llcd->llcd_ctxt->loc_imp)
310 list_move_tail(&llcd->llcd_list,
311 &lcd->lcd_llcd_list);
314 spin_unlock(&lcm->lcm_llcd_lock);
316 /* We are the only one manipulating our local list - no lock */
317 list_for_each_entry_safe(llcd,n, &lcd->lcd_llcd_list,llcd_list){
318 char *bufs[1] = {(char *)llcd->llcd_cookies};
320 list_del(&llcd->llcd_list);
321 if (llcd->llcd_cookiebytes == 0) {
322 CDEBUG(D_HA, "put empty llcd %p:%p\n",
323 llcd, llcd->llcd_ctxt);
328 down(&llcd->llcd_ctxt->loc_sem);
329 if (llcd->llcd_ctxt->loc_imp == NULL) {
330 up(&llcd->llcd_ctxt->loc_sem);
331 CWARN("import will be destroyed, put "
332 "llcd %p:%p\n", llcd, llcd->llcd_ctxt);
337 up(&llcd->llcd_ctxt->loc_sem);
338 request = ptlrpc_prep_req(import, LUSTRE_LOG_VERSION,
340 &llcd->llcd_cookiebytes,
343 if (request == NULL) {
345 CERROR("error preparing commit: rc %d\n", rc);
347 spin_lock(&lcm->lcm_llcd_lock);
348 list_splice(&lcd->lcd_llcd_list,
349 &lcm->lcm_llcd_resend);
350 INIT_LIST_HEAD(&lcd->lcd_llcd_list);
351 spin_unlock(&lcm->lcm_llcd_lock);
355 request->rq_replen = lustre_msg_size(0, NULL);
356 down(&llcd->llcd_ctxt->loc_sem);
357 if (llcd->llcd_ctxt->loc_imp == NULL) {
358 up(&llcd->llcd_ctxt->loc_sem);
359 CWARN("import will be destroyed, put "
360 "llcd %p:%p\n", llcd, llcd->llcd_ctxt);
362 ptlrpc_req_finished(request);
365 up(&llcd->llcd_ctxt->loc_sem);
366 rc = ptlrpc_queue_wait(request);
367 ptlrpc_req_finished(request);
369 /* If the RPC failed, we put this and the remaining
370 * messages onto the resend list for another time. */
376 #if 0 /* FIXME just put llcd, not put it on resend list */
377 spin_lock(&lcm->lcm_llcd_lock);
378 list_splice(&lcd->lcd_llcd_list, &lcm->lcm_llcd_resend);
379 if (++llcd->llcd_tries < 5) {
380 CERROR("commit %p failed on attempt %d: rc %d\n",
381 llcd, llcd->llcd_tries, rc);
383 list_add_tail(&llcd->llcd_list,
384 &lcm->lcm_llcd_resend);
385 spin_unlock(&lcm->lcm_llcd_lock);
387 spin_unlock(&lcm->lcm_llcd_lock);
389 CERROR("commit %p:%p drop %d cookies: rc %d\n",
390 llcd, llcd->llcd_ctxt,
391 (int)(llcd->llcd_cookiebytes /
392 sizeof(*llcd->llcd_cookies)), rc);
401 sending_list = &lcm->lcm_llcd_resend;
402 if (!list_empty(sending_list))
407 /* If we are force exiting, just drop all of the cookies. */
408 if (lcm->lcm_flags & LLOG_LCM_FL_EXIT_FORCE) {
409 spin_lock(&lcm->lcm_llcd_lock);
410 list_splice(&lcm->lcm_llcd_pending, &lcd->lcd_llcd_list);
411 list_splice(&lcm->lcm_llcd_resend, &lcd->lcd_llcd_list);
412 list_splice(&lcm->lcm_llcd_free, &lcd->lcd_llcd_list);
413 spin_unlock(&lcm->lcm_llcd_lock);
415 list_for_each_entry_safe(llcd, n, &lcd->lcd_llcd_list,llcd_list)
419 spin_lock(&lcm->lcm_thread_lock);
420 list_del(&lcd->lcd_lcm_list);
421 spin_unlock(&lcm->lcm_thread_lock);
422 OBD_FREE(lcd, sizeof(*lcd));
424 spin_lock(&lcm->lcm_thread_lock);
425 atomic_dec(&lcm->lcm_thread_total);
426 spin_unlock(&lcm->lcm_thread_lock);
427 wake_up(&lcm->lcm_waitq);
429 CDEBUG(D_HA, "%s exiting\n", current->comm);
433 int llog_start_commit_thread(void)
438 if (atomic_read(&lcm->lcm_thread_total) >= lcm->lcm_thread_max)
441 rc = kernel_thread(log_commit_thread, lcm, CLONE_VM | CLONE_FILES);
443 CERROR("error starting thread #%d: %d\n",
444 atomic_read(&lcm->lcm_thread_total), rc);
450 EXPORT_SYMBOL(llog_start_commit_thread);
452 static struct llog_process_args {
453 struct semaphore llpa_sem;
454 struct llog_ctxt *llpa_ctxt;
459 int llog_init_commit_master(void)
461 INIT_LIST_HEAD(&lcm->lcm_thread_busy);
462 INIT_LIST_HEAD(&lcm->lcm_thread_idle);
463 spin_lock_init(&lcm->lcm_thread_lock);
464 atomic_set(&lcm->lcm_thread_numidle, 0);
465 init_waitqueue_head(&lcm->lcm_waitq);
466 INIT_LIST_HEAD(&lcm->lcm_llcd_pending);
467 INIT_LIST_HEAD(&lcm->lcm_llcd_resend);
468 INIT_LIST_HEAD(&lcm->lcm_llcd_free);
469 spin_lock_init(&lcm->lcm_llcd_lock);
470 atomic_set(&lcm->lcm_llcd_numfree, 0);
471 lcm->lcm_llcd_minfree = 0;
472 lcm->lcm_thread_max = 5;
473 /* FIXME initialize semaphore for llog_process_args */
474 sema_init(&llpa.llpa_sem, 1);
478 int llog_cleanup_commit_master(int force)
480 lcm->lcm_flags |= LLOG_LCM_FL_EXIT;
482 lcm->lcm_flags |= LLOG_LCM_FL_EXIT_FORCE;
483 wake_up(&lcm->lcm_waitq);
485 wait_event_interruptible(lcm->lcm_waitq,
486 atomic_read(&lcm->lcm_thread_total) == 0);
489 EXPORT_SYMBOL(llog_cleanup_commit_master);
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 THREAD_NAME(current->comm, sizeof(current->comm) - 1, "llog_process");
507 SIGNAL_MASK_LOCK(current, flags);
508 sigfillset(¤t->blocked);
510 SIGNAL_MASK_UNLOCK(current, flags);
513 rc = llog_open(ctxt, &llh, &logid, NULL, 0);
515 CERROR("llog_create failed %d\n", rc);
518 rc = llog_init_handle(llh, LLOG_F_IS_CAT, NULL);
520 CERROR("llog_init_handle failed %d\n", rc);
525 rc = llog_cat_process(llh, (llog_cb_t)cb, NULL);
526 if (rc != LLOG_PROC_BREAK)
527 CERROR("llog_cat_process failed %d\n", rc);
529 CWARN("no callback function for recovery\n");
532 CDEBUG(D_HA, "send llcd %p:%p forcibly after recovery\n",
533 ctxt->loc_llcd, ctxt);
534 llog_sync(ctxt, NULL);
536 rc = llog_cat_put(llh);
538 CERROR("llog_cat_put failed %d\n", rc);
543 static int llog_recovery_generic(struct llog_ctxt *ctxt, void *handle,void *arg)
548 down(&llpa.llpa_sem);
549 llpa.llpa_ctxt = ctxt;
550 llpa.llpa_cb = handle;
553 rc = kernel_thread(log_process_thread, &llpa, CLONE_VM | CLONE_FILES);
555 CERROR("error starting log_process_thread: %d\n", rc);
557 CDEBUG(D_HA, "log_process_thread: %d\n", rc);
564 int llog_repl_connect(struct llog_ctxt *ctxt, int count,
565 struct llog_logid *logid, struct llog_gen *gen,
566 struct obd_uuid *uuid)
568 struct llog_canceld_ctxt *llcd;
572 /* send back llcd before recovery from llog */
573 if (ctxt->loc_llcd != NULL) {
574 CWARN("llcd %p:%p not empty\n", ctxt->loc_llcd, ctxt);
575 llog_sync(ctxt, NULL);
578 down(&ctxt->loc_sem);
579 ctxt->loc_gen = *gen;
582 CERROR("couldn't get an llcd\n");
586 llcd->llcd_ctxt = ctxt;
587 ctxt->loc_llcd = llcd;
590 rc = llog_recovery_generic(ctxt, ctxt->llog_proc_cb, logid);
592 CERROR("error recovery process: %d\n", rc);
596 EXPORT_SYMBOL(llog_repl_connect);
598 #else /* !__KERNEL__ */
600 int llog_obd_repl_cancel(struct llog_ctxt *ctxt, int count,
601 struct llog_cookie *cookies, int flags, void *data)