xref: /openbmc/linux/fs/ocfs2/cluster/quorum.c (revision 96de0e252cedffad61b3cb5e05662c591898e69a)
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  *
3  * vim: noexpandtab sw=8 ts=8 sts=0:
4  *
5  * Copyright (C) 2005 Oracle.  All rights reserved.
6  *
7  * This program is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU General Public
9  * License as published by the Free Software Foundation; either
10  * version 2 of the License, or (at your option) any later version.
11  *
12  * This program is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License for more details.
16  *
17  * You should have received a copy of the GNU General Public
18  * License along with this program; if not, write to the
19  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
20  * Boston, MA 021110-1307, USA.
21  */
22 
23 /* This quorum hack is only here until we transition to some more rational
24  * approach that is driven from userspace.  Honest.  No foolin'.
25  *
26  * Imagine two nodes lose network connectivity to each other but they're still
27  * up and operating in every other way.  Presumably a network timeout indicates
28  * that a node is broken and should be recovered.  They can't both recover each
29  * other and both carry on without serialising their access to the file system.
30  * They need to decide who is authoritative.  Now extend that problem to
31  * arbitrary groups of nodes losing connectivity between each other.
32  *
33  * So we declare that a node which has given up on connecting to a majority
34  * of nodes who are still heartbeating will fence itself.
35  *
36  * There are huge opportunities for races here.  After we give up on a node's
37  * connection we need to wait long enough to give heartbeat an opportunity
38  * to declare the node as truly dead.  We also need to be careful with the
39  * race between when we see a node start heartbeating and when we connect
40  * to it.
41  *
42  * So nodes that are in this transtion put a hold on the quorum decision
43  * with a counter.  As they fall out of this transition they drop the count
44  * and if they're the last, they fire off the decision.
45  */
46 #include <linux/kernel.h>
47 #include <linux/slab.h>
48 #include <linux/workqueue.h>
49 #include <linux/reboot.h>
50 
51 #include "heartbeat.h"
52 #include "nodemanager.h"
53 #define MLOG_MASK_PREFIX ML_QUORUM
54 #include "masklog.h"
55 #include "quorum.h"
56 
57 static struct o2quo_state {
58 	spinlock_t		qs_lock;
59 	struct work_struct	qs_work;
60 	int			qs_pending;
61 	int			qs_heartbeating;
62 	unsigned long		qs_hb_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
63 	int			qs_connected;
64 	unsigned long		qs_conn_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
65 	int			qs_holds;
66 	unsigned long		qs_hold_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
67 } o2quo_state;
68 
69 /* this is horribly heavy-handed.  It should instead flip the file
70  * system RO and call some userspace script. */
71 static void o2quo_fence_self(void)
72 {
73 	/* panic spins with interrupts enabled.  with preempt
74 	 * threads can still schedule, etc, etc */
75 	o2hb_stop_all_regions();
76 
77 	printk("ocfs2 is very sorry to be fencing this system by restarting\n");
78 	emergency_restart();
79 }
80 
81 /* Indicate that a timeout occured on a hearbeat region write. The
82  * other nodes in the cluster may consider us dead at that time so we
83  * want to "fence" ourselves so that we don't scribble on the disk
84  * after they think they've recovered us. This can't solve all
85  * problems related to writeout after recovery but this hack can at
86  * least close some of those gaps. When we have real fencing, this can
87  * go away as our node would be fenced externally before other nodes
88  * begin recovery. */
89 void o2quo_disk_timeout(void)
90 {
91 	o2quo_fence_self();
92 }
93 
94 static void o2quo_make_decision(struct work_struct *work)
95 {
96 	int quorum;
97 	int lowest_hb, lowest_reachable = 0, fence = 0;
98 	struct o2quo_state *qs = &o2quo_state;
99 
100 	spin_lock(&qs->qs_lock);
101 
102 	lowest_hb = find_first_bit(qs->qs_hb_bm, O2NM_MAX_NODES);
103 	if (lowest_hb != O2NM_MAX_NODES)
104 		lowest_reachable = test_bit(lowest_hb, qs->qs_conn_bm);
105 
106 	mlog(0, "heartbeating: %d, connected: %d, "
107 	     "lowest: %d (%sreachable)\n", qs->qs_heartbeating,
108 	     qs->qs_connected, lowest_hb, lowest_reachable ? "" : "un");
109 
110 	if (!test_bit(o2nm_this_node(), qs->qs_hb_bm) ||
111 	    qs->qs_heartbeating == 1)
112 		goto out;
113 
114 	if (qs->qs_heartbeating & 1) {
115 		/* the odd numbered cluster case is straight forward --
116 		 * if we can't talk to the majority we're hosed */
117 		quorum = (qs->qs_heartbeating + 1)/2;
118 		if (qs->qs_connected < quorum) {
119 			mlog(ML_ERROR, "fencing this node because it is "
120 			     "only connected to %u nodes and %u is needed "
121 			     "to make a quorum out of %u heartbeating nodes\n",
122 			     qs->qs_connected, quorum,
123 			     qs->qs_heartbeating);
124 			fence = 1;
125 		}
126 	} else {
127 		/* the even numbered cluster adds the possibility of each half
128 		 * of the cluster being able to talk amongst themselves.. in
129 		 * that case we're hosed if we can't talk to the group that has
130 		 * the lowest numbered node */
131 		quorum = qs->qs_heartbeating / 2;
132 		if (qs->qs_connected < quorum) {
133 			mlog(ML_ERROR, "fencing this node because it is "
134 			     "only connected to %u nodes and %u is needed "
135 			     "to make a quorum out of %u heartbeating nodes\n",
136 			     qs->qs_connected, quorum,
137 			     qs->qs_heartbeating);
138 			fence = 1;
139 		}
140 		else if ((qs->qs_connected == quorum) &&
141 			 !lowest_reachable) {
142 			mlog(ML_ERROR, "fencing this node because it is "
143 			     "connected to a half-quorum of %u out of %u "
144 			     "nodes which doesn't include the lowest active "
145 			     "node %u\n", quorum, qs->qs_heartbeating,
146 			     lowest_hb);
147 			fence = 1;
148 		}
149 	}
150 
151 out:
152 	spin_unlock(&qs->qs_lock);
153 	if (fence)
154 		o2quo_fence_self();
155 }
156 
157 static void o2quo_set_hold(struct o2quo_state *qs, u8 node)
158 {
159 	assert_spin_locked(&qs->qs_lock);
160 
161 	if (!test_and_set_bit(node, qs->qs_hold_bm)) {
162 		qs->qs_holds++;
163 		mlog_bug_on_msg(qs->qs_holds == O2NM_MAX_NODES,
164 			        "node %u\n", node);
165 		mlog(0, "node %u, %d total\n", node, qs->qs_holds);
166 	}
167 }
168 
169 static void o2quo_clear_hold(struct o2quo_state *qs, u8 node)
170 {
171 	assert_spin_locked(&qs->qs_lock);
172 
173 	if (test_and_clear_bit(node, qs->qs_hold_bm)) {
174 		mlog(0, "node %u, %d total\n", node, qs->qs_holds - 1);
175 		if (--qs->qs_holds == 0) {
176 			if (qs->qs_pending) {
177 				qs->qs_pending = 0;
178 				schedule_work(&qs->qs_work);
179 			}
180 		}
181 		mlog_bug_on_msg(qs->qs_holds < 0, "node %u, holds %d\n",
182 				node, qs->qs_holds);
183 	}
184 }
185 
186 /* as a node comes up we delay the quorum decision until we know the fate of
187  * the connection.  the hold will be droped in conn_up or hb_down.  it might be
188  * perpetuated by con_err until hb_down.  if we already have a conn, we might
189  * be dropping a hold that conn_up got. */
190 void o2quo_hb_up(u8 node)
191 {
192 	struct o2quo_state *qs = &o2quo_state;
193 
194 	spin_lock(&qs->qs_lock);
195 
196 	qs->qs_heartbeating++;
197 	mlog_bug_on_msg(qs->qs_heartbeating == O2NM_MAX_NODES,
198 		        "node %u\n", node);
199 	mlog_bug_on_msg(test_bit(node, qs->qs_hb_bm), "node %u\n", node);
200 	set_bit(node, qs->qs_hb_bm);
201 
202 	mlog(0, "node %u, %d total\n", node, qs->qs_heartbeating);
203 
204 	if (!test_bit(node, qs->qs_conn_bm))
205 		o2quo_set_hold(qs, node);
206 	else
207 		o2quo_clear_hold(qs, node);
208 
209 	spin_unlock(&qs->qs_lock);
210 }
211 
212 /* hb going down releases any holds we might have had due to this node from
213  * conn_up, conn_err, or hb_up */
214 void o2quo_hb_down(u8 node)
215 {
216 	struct o2quo_state *qs = &o2quo_state;
217 
218 	spin_lock(&qs->qs_lock);
219 
220 	qs->qs_heartbeating--;
221 	mlog_bug_on_msg(qs->qs_heartbeating < 0,
222 			"node %u, %d heartbeating\n",
223 			node, qs->qs_heartbeating);
224 	mlog_bug_on_msg(!test_bit(node, qs->qs_hb_bm), "node %u\n", node);
225 	clear_bit(node, qs->qs_hb_bm);
226 
227 	mlog(0, "node %u, %d total\n", node, qs->qs_heartbeating);
228 
229 	o2quo_clear_hold(qs, node);
230 
231 	spin_unlock(&qs->qs_lock);
232 }
233 
234 /* this tells us that we've decided that the node is still heartbeating
235  * even though we've lost it's conn.  it must only be called after conn_err
236  * and indicates that we must now make a quorum decision in the future,
237  * though we might be doing so after waiting for holds to drain.  Here
238  * we'll be dropping the hold from conn_err. */
239 void o2quo_hb_still_up(u8 node)
240 {
241 	struct o2quo_state *qs = &o2quo_state;
242 
243 	spin_lock(&qs->qs_lock);
244 
245 	mlog(0, "node %u\n", node);
246 
247 	qs->qs_pending = 1;
248 	o2quo_clear_hold(qs, node);
249 
250 	spin_unlock(&qs->qs_lock);
251 }
252 
253 /* This is analagous to hb_up.  as a node's connection comes up we delay the
254  * quorum decision until we see it heartbeating.  the hold will be droped in
255  * hb_up or hb_down.  it might be perpetuated by con_err until hb_down.  if
256  * it's already heartbeating we we might be dropping a hold that conn_up got.
257  * */
258 void o2quo_conn_up(u8 node)
259 {
260 	struct o2quo_state *qs = &o2quo_state;
261 
262 	spin_lock(&qs->qs_lock);
263 
264 	qs->qs_connected++;
265 	mlog_bug_on_msg(qs->qs_connected == O2NM_MAX_NODES,
266 		        "node %u\n", node);
267 	mlog_bug_on_msg(test_bit(node, qs->qs_conn_bm), "node %u\n", node);
268 	set_bit(node, qs->qs_conn_bm);
269 
270 	mlog(0, "node %u, %d total\n", node, qs->qs_connected);
271 
272 	if (!test_bit(node, qs->qs_hb_bm))
273 		o2quo_set_hold(qs, node);
274 	else
275 		o2quo_clear_hold(qs, node);
276 
277 	spin_unlock(&qs->qs_lock);
278 }
279 
280 /* we've decided that we won't ever be connecting to the node again.  if it's
281  * still heartbeating we grab a hold that will delay decisions until either the
282  * node stops heartbeating from hb_down or the caller decides that the node is
283  * still up and calls still_up */
284 void o2quo_conn_err(u8 node)
285 {
286 	struct o2quo_state *qs = &o2quo_state;
287 
288 	spin_lock(&qs->qs_lock);
289 
290 	if (test_bit(node, qs->qs_conn_bm)) {
291 		qs->qs_connected--;
292 		mlog_bug_on_msg(qs->qs_connected < 0,
293 				"node %u, connected %d\n",
294 				node, qs->qs_connected);
295 
296 		clear_bit(node, qs->qs_conn_bm);
297 	}
298 
299 	mlog(0, "node %u, %d total\n", node, qs->qs_connected);
300 
301 	if (test_bit(node, qs->qs_hb_bm))
302 		o2quo_set_hold(qs, node);
303 
304 	spin_unlock(&qs->qs_lock);
305 }
306 
307 void o2quo_init(void)
308 {
309 	struct o2quo_state *qs = &o2quo_state;
310 
311 	spin_lock_init(&qs->qs_lock);
312 	INIT_WORK(&qs->qs_work, o2quo_make_decision);
313 }
314 
315 void o2quo_exit(void)
316 {
317 	flush_scheduled_work();
318 }
319