4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; if not, write to the
18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19 * Boston, MA 021110-1307, USA
24 * Copyright (c) 2014, 2017, Intel Corporation.
27 * This file is part of Lustre, http://www.lustre.org/
28 * Lustre is a trademark of Sun Microsystems, Inc.
30 * lnet/lnet/net_fault.c
32 * Lustre network fault simulation
34 * Author: liang.zhen@intel.com
37 #define DEBUG_SUBSYSTEM S_LNET
39 #include <linux/random.h>
40 #include <lnet/lib-lnet.h>
41 #include <uapi/linux/lnet/lnetctl.h>
43 #define LNET_MSG_MASK (LNET_PUT_BIT | LNET_ACK_BIT | \
44 LNET_GET_BIT | LNET_REPLY_BIT)
46 struct lnet_drop_rule {
47 /** link chain on the_lnet.ln_drop_rules */
48 struct list_head dr_link;
49 /** attributes of this rule */
50 struct lnet_fault_attr dr_attr;
51 /** lock to protect \a dr_drop_at and \a dr_stat */
54 * the message sequence to drop, which means message is dropped when
55 * dr_stat.drs_count == dr_drop_at
57 unsigned long dr_drop_at;
59 * seconds to drop the next message, it's exclusive with dr_drop_at
61 time64_t dr_drop_time;
62 /** baseline to caculate dr_drop_time */
63 time64_t dr_time_base;
64 /** statistic of dropped messages */
65 struct lnet_fault_stat dr_stat;
69 lnet_fault_nid_match(lnet_nid_t nid, lnet_nid_t msg_nid)
71 if (nid == msg_nid || nid == LNET_NID_ANY)
74 if (LNET_NIDNET(nid) != LNET_NIDNET(msg_nid))
77 /* 255.255.255.255@net is wildcard for all addresses in a network */
78 return LNET_NIDADDR(nid) == LNET_NIDADDR(LNET_NID_ANY);
82 lnet_fault_attr_match(struct lnet_fault_attr *attr, lnet_nid_t src,
83 lnet_nid_t local_nid, lnet_nid_t dst,
84 unsigned int type, unsigned int portal)
86 if (!lnet_fault_nid_match(attr->fa_src, src) ||
87 !lnet_fault_nid_match(attr->fa_dst, dst) ||
88 !lnet_fault_nid_match(attr->fa_local_nid, local_nid))
91 if (!(attr->fa_msg_mask & (1 << type)))
94 /* NB: ACK and REPLY have no portal, but they should have been
95 * rejected by message mask */
96 if (attr->fa_ptl_mask != 0 && /* has portal filter */
97 !(attr->fa_ptl_mask & (1ULL << portal)))
104 lnet_fault_attr_validate(struct lnet_fault_attr *attr)
106 if (attr->fa_msg_mask == 0)
107 attr->fa_msg_mask = LNET_MSG_MASK; /* all message types */
109 if (attr->fa_ptl_mask == 0) /* no portal filter */
112 /* NB: only PUT and GET can be filtered if portal filter has been set */
113 attr->fa_msg_mask &= LNET_GET_BIT | LNET_PUT_BIT;
114 if (attr->fa_msg_mask == 0) {
115 CDEBUG(D_NET, "can't find valid message type bits %x\n",
123 lnet_fault_stat_inc(struct lnet_fault_stat *stat, unsigned int type)
125 /* NB: fs_counter is NOT updated by this function */
143 * LNet message drop simulation
147 * Add a new drop rule to LNet
148 * There is no check for duplicated drop rule, all rules will be checked for
152 lnet_drop_rule_add(struct lnet_fault_attr *attr)
154 struct lnet_drop_rule *rule;
157 if (!((attr->u.drop.da_rate == 0) ^ (attr->u.drop.da_interval == 0))) {
159 "please provide either drop rate or drop interval, "
160 "but not both at the same time %d/%d\n",
161 attr->u.drop.da_rate, attr->u.drop.da_interval);
165 if (lnet_fault_attr_validate(attr) != 0)
172 spin_lock_init(&rule->dr_lock);
174 rule->dr_attr = *attr;
175 if (attr->u.drop.da_interval != 0) {
176 rule->dr_time_base = ktime_get_seconds() + attr->u.drop.da_interval;
177 rule->dr_drop_time = ktime_get_seconds() +
178 prandom_u32_max(attr->u.drop.da_interval);
180 rule->dr_drop_at = prandom_u32_max(attr->u.drop.da_rate);
183 lnet_net_lock(LNET_LOCK_EX);
184 list_add(&rule->dr_link, &the_lnet.ln_drop_rules);
185 lnet_net_unlock(LNET_LOCK_EX);
187 CDEBUG(D_NET, "Added drop rule: src %s, dst %s, rate %d, interval %d\n",
188 libcfs_nid2str(attr->fa_src), libcfs_nid2str(attr->fa_src),
189 attr->u.drop.da_rate, attr->u.drop.da_interval);
194 * Remove matched drop rules from lnet, all rules that can match \a src and
195 * \a dst will be removed.
196 * If \a src is zero, then all rules have \a dst as destination will be remove
197 * If \a dst is zero, then all rules have \a src as source will be removed
198 * If both of them are zero, all rules will be removed
201 lnet_drop_rule_del(lnet_nid_t src, lnet_nid_t dst)
203 struct lnet_drop_rule *rule;
204 struct lnet_drop_rule *tmp;
209 lnet_net_lock(LNET_LOCK_EX);
210 list_for_each_entry_safe(rule, tmp, &the_lnet.ln_drop_rules, dr_link) {
211 if (rule->dr_attr.fa_src != src && src != 0)
214 if (rule->dr_attr.fa_dst != dst && dst != 0)
217 list_move(&rule->dr_link, &zombies);
219 lnet_net_unlock(LNET_LOCK_EX);
221 list_for_each_entry_safe(rule, tmp, &zombies, dr_link) {
222 CDEBUG(D_NET, "Remove drop rule: src %s->dst: %s (1/%d, %d)\n",
223 libcfs_nid2str(rule->dr_attr.fa_src),
224 libcfs_nid2str(rule->dr_attr.fa_dst),
225 rule->dr_attr.u.drop.da_rate,
226 rule->dr_attr.u.drop.da_interval);
228 list_del(&rule->dr_link);
237 * List drop rule at position of \a pos
240 lnet_drop_rule_list(int pos, struct lnet_fault_attr *attr,
241 struct lnet_fault_stat *stat)
243 struct lnet_drop_rule *rule;
249 cpt = lnet_net_lock_current();
250 list_for_each_entry(rule, &the_lnet.ln_drop_rules, dr_link) {
254 spin_lock(&rule->dr_lock);
255 *attr = rule->dr_attr;
256 *stat = rule->dr_stat;
257 spin_unlock(&rule->dr_lock);
262 lnet_net_unlock(cpt);
267 * reset counters for all drop rules
270 lnet_drop_rule_reset(void)
272 struct lnet_drop_rule *rule;
276 cpt = lnet_net_lock_current();
278 list_for_each_entry(rule, &the_lnet.ln_drop_rules, dr_link) {
279 struct lnet_fault_attr *attr = &rule->dr_attr;
281 spin_lock(&rule->dr_lock);
283 memset(&rule->dr_stat, 0, sizeof(rule->dr_stat));
284 if (attr->u.drop.da_rate != 0) {
285 rule->dr_drop_at = prandom_u32_max(attr->u.drop.da_rate);
287 rule->dr_drop_time = ktime_get_seconds() +
288 prandom_u32_max(attr->u.drop.da_interval);
289 rule->dr_time_base = ktime_get_seconds() + attr->u.drop.da_interval;
291 spin_unlock(&rule->dr_lock);
294 lnet_net_unlock(cpt);
299 lnet_fault_match_health(enum lnet_msg_hstatus *hstatus, __u32 mask)
306 /* assign a random failure */
307 choice = prandom_u32_max(LNET_MSG_STATUS_END - LNET_MSG_STATUS_OK);
311 if (mask == HSTATUS_RANDOM) {
316 if (mask & (1 << choice)) {
321 /* round to the closest ON bit */
323 best_delta = HSTATUS_END;
325 if (mask & (1 << i)) {
329 if (delta < best_delta) {
341 * check source/destination NID, portal, message type and drop rate,
342 * decide whether should drop this message or not
345 drop_rule_match(struct lnet_drop_rule *rule, lnet_nid_t src,
346 lnet_nid_t local_nid, lnet_nid_t dst,
347 unsigned int type, unsigned int portal,
348 enum lnet_msg_hstatus *hstatus)
350 struct lnet_fault_attr *attr = &rule->dr_attr;
353 if (!lnet_fault_attr_match(attr, src, local_nid, dst, type, portal))
356 if (attr->u.drop.da_drop_all) {
357 CDEBUG(D_NET, "set to drop all messages\n");
363 * if we're trying to match a health status error but it hasn't
364 * been set in the rule, then don't match
366 if ((hstatus && !attr->u.drop.da_health_error_mask) ||
367 (!hstatus && attr->u.drop.da_health_error_mask))
370 /* match this rule, check drop rate now */
371 spin_lock(&rule->dr_lock);
372 if (attr->u.drop.da_random) {
373 int value = prandom_u32_max(attr->u.drop.da_interval);
374 if (value >= (attr->u.drop.da_interval / 2))
378 } else if (rule->dr_drop_time != 0) { /* time based drop */
379 time64_t now = ktime_get_seconds();
381 rule->dr_stat.fs_count++;
382 drop = now >= rule->dr_drop_time;
384 if (now > rule->dr_time_base)
385 rule->dr_time_base = now;
387 rule->dr_drop_time = rule->dr_time_base +
388 prandom_u32_max(attr->u.drop.da_interval);
389 rule->dr_time_base += attr->u.drop.da_interval;
391 CDEBUG(D_NET, "Drop Rule %s->%s: next drop : %lld\n",
392 libcfs_nid2str(attr->fa_src),
393 libcfs_nid2str(attr->fa_dst),
397 } else { /* rate based drop */
400 drop = rule->dr_stat.fs_count++ == rule->dr_drop_at;
401 count = rule->dr_stat.fs_count;
402 if (do_div(count, attr->u.drop.da_rate) == 0) {
403 rule->dr_drop_at = rule->dr_stat.fs_count +
404 prandom_u32_max(attr->u.drop.da_rate);
405 CDEBUG(D_NET, "Drop Rule %s->%s: next drop: %lu\n",
406 libcfs_nid2str(attr->fa_src),
407 libcfs_nid2str(attr->fa_dst), rule->dr_drop_at);
413 if (drop) { /* drop this message, update counters */
415 lnet_fault_match_health(hstatus,
416 attr->u.drop.da_health_error_mask);
417 lnet_fault_stat_inc(&rule->dr_stat, type);
418 rule->dr_stat.u.drop.ds_dropped++;
421 spin_unlock(&rule->dr_lock);
426 * Check if message from \a src to \a dst can match any existed drop rule
429 lnet_drop_rule_match(struct lnet_hdr *hdr,
430 lnet_nid_t local_nid,
431 enum lnet_msg_hstatus *hstatus)
433 lnet_nid_t src = le64_to_cpu(hdr->src_nid);
434 lnet_nid_t dst = le64_to_cpu(hdr->dest_nid);
435 unsigned int typ = le32_to_cpu(hdr->type);
436 struct lnet_drop_rule *rule;
437 unsigned int ptl = -1;
441 /* NB: if Portal is specified, then only PUT and GET will be
442 * filtered by drop rule */
443 if (typ == LNET_MSG_PUT)
444 ptl = le32_to_cpu(hdr->msg.put.ptl_index);
445 else if (typ == LNET_MSG_GET)
446 ptl = le32_to_cpu(hdr->msg.get.ptl_index);
448 cpt = lnet_net_lock_current();
449 list_for_each_entry(rule, &the_lnet.ln_drop_rules, dr_link) {
450 drop = drop_rule_match(rule, src, local_nid, dst, typ, ptl,
455 lnet_net_unlock(cpt);
461 * LNet Delay Simulation
463 /** timestamp (second) to send delayed message */
464 #define msg_delay_send msg_ev.hdr_data
466 struct lnet_delay_rule {
467 /** link chain on the_lnet.ln_delay_rules */
468 struct list_head dl_link;
469 /** link chain on delay_dd.dd_sched_rules */
470 struct list_head dl_sched_link;
471 /** attributes of this rule */
472 struct lnet_fault_attr dl_attr;
473 /** lock to protect \a below members */
475 /** refcount of delay rule */
476 atomic_t dl_refcount;
478 * the message sequence to delay, which means message is delayed when
479 * dl_stat.fs_count == dl_delay_at
481 unsigned long dl_delay_at;
483 * seconds to delay the next message, it's exclusive with dl_delay_at
485 time64_t dl_delay_time;
486 /** baseline to caculate dl_delay_time */
487 time64_t dl_time_base;
488 /** seconds until we send the next delayed message */
489 time64_t dl_msg_send;
490 /** delayed message list */
491 struct list_head dl_msg_list;
492 /** statistic of delayed messages */
493 struct lnet_fault_stat dl_stat;
494 /** timer to wakeup delay_daemon */
495 struct timer_list dl_timer;
498 struct delay_daemon_data {
499 /** serialise rule add/remove */
500 struct mutex dd_mutex;
501 /** protect rules on \a dd_sched_rules */
503 /** scheduled delay rules (by timer) */
504 struct list_head dd_sched_rules;
505 /** deamon thread sleeps at here */
506 wait_queue_head_t dd_waitq;
507 /** controler (lctl command) wait at here */
508 wait_queue_head_t dd_ctl_waitq;
509 /** deamon is running */
510 unsigned int dd_running;
511 /** deamon stopped */
512 unsigned int dd_stopped;
515 static struct delay_daemon_data delay_dd;
518 delay_rule_decref(struct lnet_delay_rule *rule)
520 if (atomic_dec_and_test(&rule->dl_refcount)) {
521 LASSERT(list_empty(&rule->dl_sched_link));
522 LASSERT(list_empty(&rule->dl_msg_list));
523 LASSERT(list_empty(&rule->dl_link));
530 * check source/destination NID, portal, message type and delay rate,
531 * decide whether should delay this message or not
534 delay_rule_match(struct lnet_delay_rule *rule, lnet_nid_t src,
535 lnet_nid_t dst, unsigned int type, unsigned int portal,
536 struct lnet_msg *msg)
538 struct lnet_fault_attr *attr = &rule->dl_attr;
541 if (!lnet_fault_attr_match(attr, src, LNET_NID_ANY,
545 /* match this rule, check delay rate now */
546 spin_lock(&rule->dl_lock);
547 if (rule->dl_delay_time != 0) { /* time based delay */
548 time64_t now = ktime_get_seconds();
550 rule->dl_stat.fs_count++;
551 delay = now >= rule->dl_delay_time;
553 if (now > rule->dl_time_base)
554 rule->dl_time_base = now;
556 rule->dl_delay_time = rule->dl_time_base +
557 prandom_u32_max(attr->u.delay.la_interval);
558 rule->dl_time_base += attr->u.delay.la_interval;
560 CDEBUG(D_NET, "Delay Rule %s->%s: next delay : %lld\n",
561 libcfs_nid2str(attr->fa_src),
562 libcfs_nid2str(attr->fa_dst),
563 rule->dl_delay_time);
566 } else { /* rate based delay */
569 delay = rule->dl_stat.fs_count++ == rule->dl_delay_at;
570 /* generate the next random rate sequence */
571 count = rule->dl_stat.fs_count;
572 if (do_div(count, attr->u.delay.la_rate) == 0) {
573 rule->dl_delay_at = rule->dl_stat.fs_count +
574 prandom_u32_max(attr->u.delay.la_rate);
575 CDEBUG(D_NET, "Delay Rule %s->%s: next delay: %lu\n",
576 libcfs_nid2str(attr->fa_src),
577 libcfs_nid2str(attr->fa_dst), rule->dl_delay_at);
582 spin_unlock(&rule->dl_lock);
586 /* delay this message, update counters */
587 lnet_fault_stat_inc(&rule->dl_stat, type);
588 rule->dl_stat.u.delay.ls_delayed++;
590 list_add_tail(&msg->msg_list, &rule->dl_msg_list);
591 msg->msg_delay_send = ktime_get_seconds() + attr->u.delay.la_latency;
592 if (rule->dl_msg_send == -1) {
593 rule->dl_msg_send = msg->msg_delay_send;
594 mod_timer(&rule->dl_timer,
595 jiffies + cfs_time_seconds(rule->dl_msg_send));
598 spin_unlock(&rule->dl_lock);
603 * check if \a msg can match any Delay Rule, receiving of this message
604 * will be delayed if there is a match.
607 lnet_delay_rule_match_locked(struct lnet_hdr *hdr, struct lnet_msg *msg)
609 struct lnet_delay_rule *rule;
610 lnet_nid_t src = le64_to_cpu(hdr->src_nid);
611 lnet_nid_t dst = le64_to_cpu(hdr->dest_nid);
612 unsigned int typ = le32_to_cpu(hdr->type);
613 unsigned int ptl = -1;
615 /* NB: called with hold of lnet_net_lock */
617 /* NB: if Portal is specified, then only PUT and GET will be
618 * filtered by delay rule */
619 if (typ == LNET_MSG_PUT)
620 ptl = le32_to_cpu(hdr->msg.put.ptl_index);
621 else if (typ == LNET_MSG_GET)
622 ptl = le32_to_cpu(hdr->msg.get.ptl_index);
624 list_for_each_entry(rule, &the_lnet.ln_delay_rules, dl_link) {
625 if (delay_rule_match(rule, src, dst, typ, ptl, msg))
632 /** check out delayed messages for send */
634 delayed_msg_check(struct lnet_delay_rule *rule, bool all,
635 struct list_head *msg_list)
637 struct lnet_msg *msg;
638 struct lnet_msg *tmp;
639 time64_t now = ktime_get_seconds();
641 if (!all && cfs_time_seconds(rule->dl_msg_send) > now)
644 spin_lock(&rule->dl_lock);
645 list_for_each_entry_safe(msg, tmp, &rule->dl_msg_list, msg_list) {
646 if (!all && msg->msg_delay_send > now)
649 msg->msg_delay_send = 0;
650 list_move_tail(&msg->msg_list, msg_list);
653 if (list_empty(&rule->dl_msg_list)) {
654 del_timer(&rule->dl_timer);
655 rule->dl_msg_send = -1;
657 } else if (!list_empty(msg_list)) {
658 /* dequeued some timedout messages, update timer for the
659 * next delayed message on rule */
660 msg = list_entry(rule->dl_msg_list.next,
661 struct lnet_msg, msg_list);
662 rule->dl_msg_send = msg->msg_delay_send;
663 mod_timer(&rule->dl_timer,
664 jiffies + cfs_time_seconds(rule->dl_msg_send));
666 spin_unlock(&rule->dl_lock);
670 delayed_msg_process(struct list_head *msg_list, bool drop)
672 struct lnet_msg *msg;
674 while (!list_empty(msg_list)) {
679 msg = list_entry(msg_list->next, struct lnet_msg, msg_list);
680 LASSERT(msg->msg_rxpeer != NULL);
681 LASSERT(msg->msg_rxni != NULL);
684 cpt = msg->msg_rx_cpt;
686 list_del_init(&msg->msg_list);
690 } else if (!msg->msg_routing) {
691 rc = lnet_parse_local(ni, msg);
697 rc = lnet_parse_forward_locked(ni, msg);
698 lnet_net_unlock(cpt);
702 lnet_ni_recv(ni, msg->msg_private, msg, 0,
703 0, msg->msg_len, msg->msg_len);
705 case LNET_CREDIT_WAIT:
707 default: /* failures */
712 lnet_drop_message(ni, cpt, msg->msg_private, msg->msg_len,
714 lnet_finalize(msg, rc);
719 * Process delayed messages for scheduled rules
720 * This function can either be called by delay_rule_daemon, or by lnet_finalise
723 lnet_delay_rule_check(void)
725 struct lnet_delay_rule *rule;
729 if (list_empty(&delay_dd.dd_sched_rules))
732 spin_lock_bh(&delay_dd.dd_lock);
733 if (list_empty(&delay_dd.dd_sched_rules)) {
734 spin_unlock_bh(&delay_dd.dd_lock);
738 rule = list_entry(delay_dd.dd_sched_rules.next,
739 struct lnet_delay_rule, dl_sched_link);
740 list_del_init(&rule->dl_sched_link);
741 spin_unlock_bh(&delay_dd.dd_lock);
743 delayed_msg_check(rule, false, &msgs);
744 delay_rule_decref(rule); /* -1 for delay_dd.dd_sched_rules */
747 if (!list_empty(&msgs))
748 delayed_msg_process(&msgs, false);
751 /** deamon thread to handle delayed messages */
753 lnet_delay_rule_daemon(void *arg)
755 delay_dd.dd_running = 1;
756 wake_up(&delay_dd.dd_ctl_waitq);
758 while (delay_dd.dd_running) {
759 wait_event_interruptible(delay_dd.dd_waitq,
760 !delay_dd.dd_running ||
761 !list_empty(&delay_dd.dd_sched_rules));
762 lnet_delay_rule_check();
765 /* in case more rules have been enqueued after my last check */
766 lnet_delay_rule_check();
767 delay_dd.dd_stopped = 1;
768 wake_up(&delay_dd.dd_ctl_waitq);
774 delay_timer_cb(cfs_timer_cb_arg_t data)
776 struct lnet_delay_rule *rule = cfs_from_timer(rule, data, dl_timer);
778 spin_lock_bh(&delay_dd.dd_lock);
779 if (list_empty(&rule->dl_sched_link) && delay_dd.dd_running) {
780 atomic_inc(&rule->dl_refcount);
781 list_add_tail(&rule->dl_sched_link, &delay_dd.dd_sched_rules);
782 wake_up(&delay_dd.dd_waitq);
784 spin_unlock_bh(&delay_dd.dd_lock);
788 * Add a new delay rule to LNet
789 * There is no check for duplicated delay rule, all rules will be checked for
793 lnet_delay_rule_add(struct lnet_fault_attr *attr)
795 struct lnet_delay_rule *rule;
799 if (!((attr->u.delay.la_rate == 0) ^
800 (attr->u.delay.la_interval == 0))) {
802 "please provide either delay rate or delay interval, "
803 "but not both at the same time %d/%d\n",
804 attr->u.delay.la_rate, attr->u.delay.la_interval);
808 if (attr->u.delay.la_latency == 0) {
809 CDEBUG(D_NET, "delay latency cannot be zero\n");
813 if (lnet_fault_attr_validate(attr) != 0)
820 mutex_lock(&delay_dd.dd_mutex);
821 if (!delay_dd.dd_running) {
822 struct task_struct *task;
824 /* NB: although LND threads will process delayed message
825 * in lnet_finalize, but there is no guarantee that LND
826 * threads will be waken up if no other message needs to
828 * Only one daemon thread, performance is not the concern
829 * of this simualation module.
831 task = kthread_run(lnet_delay_rule_daemon, NULL, "lnet_dd");
836 wait_event(delay_dd.dd_ctl_waitq, delay_dd.dd_running);
839 cfs_timer_setup(&rule->dl_timer, delay_timer_cb,
840 (unsigned long)rule, 0);
842 spin_lock_init(&rule->dl_lock);
843 INIT_LIST_HEAD(&rule->dl_msg_list);
844 INIT_LIST_HEAD(&rule->dl_sched_link);
846 rule->dl_attr = *attr;
847 if (attr->u.delay.la_interval != 0) {
848 rule->dl_time_base = ktime_get_seconds() +
849 attr->u.delay.la_interval;
850 rule->dl_delay_time = ktime_get_seconds() +
851 prandom_u32_max(attr->u.delay.la_interval);
853 rule->dl_delay_at = prandom_u32_max(attr->u.delay.la_rate);
856 rule->dl_msg_send = -1;
858 lnet_net_lock(LNET_LOCK_EX);
859 atomic_set(&rule->dl_refcount, 1);
860 list_add(&rule->dl_link, &the_lnet.ln_delay_rules);
861 lnet_net_unlock(LNET_LOCK_EX);
863 CDEBUG(D_NET, "Added delay rule: src %s, dst %s, rate %d\n",
864 libcfs_nid2str(attr->fa_src), libcfs_nid2str(attr->fa_src),
865 attr->u.delay.la_rate);
867 mutex_unlock(&delay_dd.dd_mutex);
870 mutex_unlock(&delay_dd.dd_mutex);
876 * Remove matched Delay Rules from lnet, if \a shutdown is true or both \a src
877 * and \a dst are zero, all rules will be removed, otherwise only matched rules
879 * If \a src is zero, then all rules have \a dst as destination will be remove
880 * If \a dst is zero, then all rules have \a src as source will be removed
882 * When a delay rule is removed, all delayed messages of this rule will be
883 * processed immediately.
886 lnet_delay_rule_del(lnet_nid_t src, lnet_nid_t dst, bool shutdown)
888 struct lnet_delay_rule *rule;
889 struct lnet_delay_rule *tmp;
890 LIST_HEAD(rule_list);
899 mutex_lock(&delay_dd.dd_mutex);
900 lnet_net_lock(LNET_LOCK_EX);
902 list_for_each_entry_safe(rule, tmp, &the_lnet.ln_delay_rules, dl_link) {
903 if (rule->dl_attr.fa_src != src && src != 0)
906 if (rule->dl_attr.fa_dst != dst && dst != 0)
909 CDEBUG(D_NET, "Remove delay rule: src %s->dst: %s (1/%d, %d)\n",
910 libcfs_nid2str(rule->dl_attr.fa_src),
911 libcfs_nid2str(rule->dl_attr.fa_dst),
912 rule->dl_attr.u.delay.la_rate,
913 rule->dl_attr.u.delay.la_interval);
914 /* refcount is taken over by rule_list */
915 list_move(&rule->dl_link, &rule_list);
918 /* check if we need to shutdown delay_daemon */
919 cleanup = list_empty(&the_lnet.ln_delay_rules) &&
920 !list_empty(&rule_list);
921 lnet_net_unlock(LNET_LOCK_EX);
923 list_for_each_entry_safe(rule, tmp, &rule_list, dl_link) {
924 list_del_init(&rule->dl_link);
926 del_timer_sync(&rule->dl_timer);
927 delayed_msg_check(rule, true, &msg_list);
928 delay_rule_decref(rule); /* -1 for the_lnet.ln_delay_rules */
932 if (cleanup) { /* no more delay rule, shutdown delay_daemon */
933 LASSERT(delay_dd.dd_running);
934 delay_dd.dd_running = 0;
935 wake_up(&delay_dd.dd_waitq);
937 while (!delay_dd.dd_stopped)
938 wait_event(delay_dd.dd_ctl_waitq, delay_dd.dd_stopped);
940 mutex_unlock(&delay_dd.dd_mutex);
942 if (!list_empty(&msg_list))
943 delayed_msg_process(&msg_list, shutdown);
949 * List Delay Rule at position of \a pos
952 lnet_delay_rule_list(int pos, struct lnet_fault_attr *attr,
953 struct lnet_fault_stat *stat)
955 struct lnet_delay_rule *rule;
961 cpt = lnet_net_lock_current();
962 list_for_each_entry(rule, &the_lnet.ln_delay_rules, dl_link) {
966 spin_lock(&rule->dl_lock);
967 *attr = rule->dl_attr;
968 *stat = rule->dl_stat;
969 spin_unlock(&rule->dl_lock);
974 lnet_net_unlock(cpt);
979 * reset counters for all Delay Rules
982 lnet_delay_rule_reset(void)
984 struct lnet_delay_rule *rule;
988 cpt = lnet_net_lock_current();
990 list_for_each_entry(rule, &the_lnet.ln_delay_rules, dl_link) {
991 struct lnet_fault_attr *attr = &rule->dl_attr;
993 spin_lock(&rule->dl_lock);
995 memset(&rule->dl_stat, 0, sizeof(rule->dl_stat));
996 if (attr->u.delay.la_rate != 0) {
997 rule->dl_delay_at = prandom_u32_max(attr->u.delay.la_rate);
999 rule->dl_delay_time = ktime_get_seconds() +
1000 prandom_u32_max(attr->u.delay.la_interval);
1001 rule->dl_time_base = ktime_get_seconds() +
1002 attr->u.delay.la_interval;
1004 spin_unlock(&rule->dl_lock);
1007 lnet_net_unlock(cpt);
1012 lnet_fault_ctl(int opc, struct libcfs_ioctl_data *data)
1014 struct lnet_fault_attr *attr;
1015 struct lnet_fault_stat *stat;
1017 attr = (struct lnet_fault_attr *)data->ioc_inlbuf1;
1023 case LNET_CTL_DROP_ADD:
1027 return lnet_drop_rule_add(attr);
1029 case LNET_CTL_DROP_DEL:
1033 data->ioc_count = lnet_drop_rule_del(attr->fa_src,
1037 case LNET_CTL_DROP_RESET:
1038 lnet_drop_rule_reset();
1041 case LNET_CTL_DROP_LIST:
1042 stat = (struct lnet_fault_stat *)data->ioc_inlbuf2;
1043 if (attr == NULL || stat == NULL)
1046 return lnet_drop_rule_list(data->ioc_count, attr, stat);
1048 case LNET_CTL_DELAY_ADD:
1052 return lnet_delay_rule_add(attr);
1054 case LNET_CTL_DELAY_DEL:
1058 data->ioc_count = lnet_delay_rule_del(attr->fa_src,
1059 attr->fa_dst, false);
1062 case LNET_CTL_DELAY_RESET:
1063 lnet_delay_rule_reset();
1066 case LNET_CTL_DELAY_LIST:
1067 stat = (struct lnet_fault_stat *)data->ioc_inlbuf2;
1068 if (attr == NULL || stat == NULL)
1071 return lnet_delay_rule_list(data->ioc_count, attr, stat);
1076 lnet_fault_init(void)
1078 BUILD_BUG_ON(LNET_PUT_BIT != 1 << LNET_MSG_PUT);
1079 BUILD_BUG_ON(LNET_ACK_BIT != 1 << LNET_MSG_ACK);
1080 BUILD_BUG_ON(LNET_GET_BIT != 1 << LNET_MSG_GET);
1081 BUILD_BUG_ON(LNET_REPLY_BIT != 1 << LNET_MSG_REPLY);
1083 mutex_init(&delay_dd.dd_mutex);
1084 spin_lock_init(&delay_dd.dd_lock);
1085 init_waitqueue_head(&delay_dd.dd_waitq);
1086 init_waitqueue_head(&delay_dd.dd_ctl_waitq);
1087 INIT_LIST_HEAD(&delay_dd.dd_sched_rules);
1093 lnet_fault_fini(void)
1095 lnet_drop_rule_del(0, 0);
1096 lnet_delay_rule_del(0, 0, true);
1098 LASSERT(list_empty(&the_lnet.ln_drop_rules));
1099 LASSERT(list_empty(&the_lnet.ln_delay_rules));
1100 LASSERT(list_empty(&delay_dd.dd_sched_rules));