xref: /openbmc/linux/fs/ocfs2/cluster/quorum.c (revision 3fc41476)
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/workqueue.h>
48 #include <linux/reboot.h>
49 
50 #include "heartbeat.h"
51 #include "nodemanager.h"
52 #define MLOG_MASK_PREFIX ML_QUORUM
53 #include "masklog.h"
54 #include "quorum.h"
55 
56 static struct o2quo_state {
57 	spinlock_t		qs_lock;
58 	struct work_struct	qs_work;
59 	int			qs_pending;
60 	int			qs_heartbeating;
61 	unsigned long		qs_hb_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
62 	int			qs_connected;
63 	unsigned long		qs_conn_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
64 	int			qs_holds;
65 	unsigned long		qs_hold_bm[BITS_TO_LONGS(O2NM_MAX_NODES)];
66 } o2quo_state;
67 
68 /* this is horribly heavy-handed.  It should instead flip the file
69  * system RO and call some userspace script. */
70 static void o2quo_fence_self(void)
71 {
72 	/* panic spins with interrupts enabled.  with preempt
73 	 * threads can still schedule, etc, etc */
74 	o2hb_stop_all_regions();
75 
76 	switch (o2nm_single_cluster->cl_fence_method) {
77 	case O2NM_FENCE_PANIC:
78 		panic("*** ocfs2 is very sorry to be fencing this system by "
79 		      "panicing ***\n");
80 		break;
81 	default:
82 		WARN_ON(o2nm_single_cluster->cl_fence_method >=
83 			O2NM_FENCE_METHODS);
84 		/* fall through */
85 	case O2NM_FENCE_RESET:
86 		printk(KERN_ERR "*** ocfs2 is very sorry to be fencing this "
87 		       "system by restarting ***\n");
88 		emergency_restart();
89 		break;
90 	};
91 }
92 
93 /* Indicate that a timeout occurred on a hearbeat region write. The
94  * other nodes in the cluster may consider us dead at that time so we
95  * want to "fence" ourselves so that we don't scribble on the disk
96  * after they think they've recovered us. This can't solve all
97  * problems related to writeout after recovery but this hack can at
98  * least close some of those gaps. When we have real fencing, this can
99  * go away as our node would be fenced externally before other nodes
100  * begin recovery. */
101 void o2quo_disk_timeout(void)
102 {
103 	o2quo_fence_self();
104 }
105 
106 static void o2quo_make_decision(struct work_struct *work)
107 {
108 	int quorum;
109 	int lowest_hb, lowest_reachable = 0, fence = 0;
110 	struct o2quo_state *qs = &o2quo_state;
111 
112 	spin_lock(&qs->qs_lock);
113 
114 	lowest_hb = find_first_bit(qs->qs_hb_bm, O2NM_MAX_NODES);
115 	if (lowest_hb != O2NM_MAX_NODES)
116 		lowest_reachable = test_bit(lowest_hb, qs->qs_conn_bm);
117 
118 	mlog(0, "heartbeating: %d, connected: %d, "
119 	     "lowest: %d (%sreachable)\n", qs->qs_heartbeating,
120 	     qs->qs_connected, lowest_hb, lowest_reachable ? "" : "un");
121 
122 	if (!test_bit(o2nm_this_node(), qs->qs_hb_bm) ||
123 	    qs->qs_heartbeating == 1)
124 		goto out;
125 
126 	if (qs->qs_heartbeating & 1) {
127 		/* the odd numbered cluster case is straight forward --
128 		 * if we can't talk to the majority we're hosed */
129 		quorum = (qs->qs_heartbeating + 1)/2;
130 		if (qs->qs_connected < quorum) {
131 			mlog(ML_ERROR, "fencing this node because it is "
132 			     "only connected to %u nodes and %u is needed "
133 			     "to make a quorum out of %u heartbeating nodes\n",
134 			     qs->qs_connected, quorum,
135 			     qs->qs_heartbeating);
136 			fence = 1;
137 		}
138 	} else {
139 		/* the even numbered cluster adds the possibility of each half
140 		 * of the cluster being able to talk amongst themselves.. in
141 		 * that case we're hosed if we can't talk to the group that has
142 		 * the lowest numbered node */
143 		quorum = qs->qs_heartbeating / 2;
144 		if (qs->qs_connected < quorum) {
145 			mlog(ML_ERROR, "fencing this node because it is "
146 			     "only connected to %u nodes and %u is needed "
147 			     "to make a quorum out of %u heartbeating nodes\n",
148 			     qs->qs_connected, quorum,
149 			     qs->qs_heartbeating);
150 			fence = 1;
151 		}
152 		else if ((qs->qs_connected == quorum) &&
153 			 !lowest_reachable) {
154 			mlog(ML_ERROR, "fencing this node because it is "
155 			     "connected to a half-quorum of %u out of %u "
156 			     "nodes which doesn't include the lowest active "
157 			     "node %u\n", quorum, qs->qs_heartbeating,
158 			     lowest_hb);
159 			fence = 1;
160 		}
161 	}
162 
163 out:
164 	if (fence) {
165 		spin_unlock(&qs->qs_lock);
166 		o2quo_fence_self();
167 	} else {
168 		mlog(ML_NOTICE, "not fencing this node, heartbeating: %d, "
169 			"connected: %d, lowest: %d (%sreachable)\n",
170 			qs->qs_heartbeating, qs->qs_connected, lowest_hb,
171 			lowest_reachable ? "" : "un");
172 		spin_unlock(&qs->qs_lock);
173 
174 	}
175 
176 }
177 
178 static void o2quo_set_hold(struct o2quo_state *qs, u8 node)
179 {
180 	assert_spin_locked(&qs->qs_lock);
181 
182 	if (!test_and_set_bit(node, qs->qs_hold_bm)) {
183 		qs->qs_holds++;
184 		mlog_bug_on_msg(qs->qs_holds == O2NM_MAX_NODES,
185 			        "node %u\n", node);
186 		mlog(0, "node %u, %d total\n", node, qs->qs_holds);
187 	}
188 }
189 
190 static void o2quo_clear_hold(struct o2quo_state *qs, u8 node)
191 {
192 	assert_spin_locked(&qs->qs_lock);
193 
194 	if (test_and_clear_bit(node, qs->qs_hold_bm)) {
195 		mlog(0, "node %u, %d total\n", node, qs->qs_holds - 1);
196 		if (--qs->qs_holds == 0) {
197 			if (qs->qs_pending) {
198 				qs->qs_pending = 0;
199 				schedule_work(&qs->qs_work);
200 			}
201 		}
202 		mlog_bug_on_msg(qs->qs_holds < 0, "node %u, holds %d\n",
203 				node, qs->qs_holds);
204 	}
205 }
206 
207 /* as a node comes up we delay the quorum decision until we know the fate of
208  * the connection.  the hold will be droped in conn_up or hb_down.  it might be
209  * perpetuated by con_err until hb_down.  if we already have a conn, we might
210  * be dropping a hold that conn_up got. */
211 void o2quo_hb_up(u8 node)
212 {
213 	struct o2quo_state *qs = &o2quo_state;
214 
215 	spin_lock(&qs->qs_lock);
216 
217 	qs->qs_heartbeating++;
218 	mlog_bug_on_msg(qs->qs_heartbeating == O2NM_MAX_NODES,
219 		        "node %u\n", node);
220 	mlog_bug_on_msg(test_bit(node, qs->qs_hb_bm), "node %u\n", node);
221 	set_bit(node, qs->qs_hb_bm);
222 
223 	mlog(0, "node %u, %d total\n", node, qs->qs_heartbeating);
224 
225 	if (!test_bit(node, qs->qs_conn_bm))
226 		o2quo_set_hold(qs, node);
227 	else
228 		o2quo_clear_hold(qs, node);
229 
230 	spin_unlock(&qs->qs_lock);
231 }
232 
233 /* hb going down releases any holds we might have had due to this node from
234  * conn_up, conn_err, or hb_up */
235 void o2quo_hb_down(u8 node)
236 {
237 	struct o2quo_state *qs = &o2quo_state;
238 
239 	spin_lock(&qs->qs_lock);
240 
241 	qs->qs_heartbeating--;
242 	mlog_bug_on_msg(qs->qs_heartbeating < 0,
243 			"node %u, %d heartbeating\n",
244 			node, qs->qs_heartbeating);
245 	mlog_bug_on_msg(!test_bit(node, qs->qs_hb_bm), "node %u\n", node);
246 	clear_bit(node, qs->qs_hb_bm);
247 
248 	mlog(0, "node %u, %d total\n", node, qs->qs_heartbeating);
249 
250 	o2quo_clear_hold(qs, node);
251 
252 	spin_unlock(&qs->qs_lock);
253 }
254 
255 /* this tells us that we've decided that the node is still heartbeating
256  * even though we've lost it's conn.  it must only be called after conn_err
257  * and indicates that we must now make a quorum decision in the future,
258  * though we might be doing so after waiting for holds to drain.  Here
259  * we'll be dropping the hold from conn_err. */
260 void o2quo_hb_still_up(u8 node)
261 {
262 	struct o2quo_state *qs = &o2quo_state;
263 
264 	spin_lock(&qs->qs_lock);
265 
266 	mlog(0, "node %u\n", node);
267 
268 	qs->qs_pending = 1;
269 	o2quo_clear_hold(qs, node);
270 
271 	spin_unlock(&qs->qs_lock);
272 }
273 
274 /* This is analogous to hb_up.  as a node's connection comes up we delay the
275  * quorum decision until we see it heartbeating.  the hold will be droped in
276  * hb_up or hb_down.  it might be perpetuated by con_err until hb_down.  if
277  * it's already heartbeating we might be dropping a hold that conn_up got.
278  * */
279 void o2quo_conn_up(u8 node)
280 {
281 	struct o2quo_state *qs = &o2quo_state;
282 
283 	spin_lock(&qs->qs_lock);
284 
285 	qs->qs_connected++;
286 	mlog_bug_on_msg(qs->qs_connected == O2NM_MAX_NODES,
287 		        "node %u\n", node);
288 	mlog_bug_on_msg(test_bit(node, qs->qs_conn_bm), "node %u\n", node);
289 	set_bit(node, qs->qs_conn_bm);
290 
291 	mlog(0, "node %u, %d total\n", node, qs->qs_connected);
292 
293 	if (!test_bit(node, qs->qs_hb_bm))
294 		o2quo_set_hold(qs, node);
295 	else
296 		o2quo_clear_hold(qs, node);
297 
298 	spin_unlock(&qs->qs_lock);
299 }
300 
301 /* we've decided that we won't ever be connecting to the node again.  if it's
302  * still heartbeating we grab a hold that will delay decisions until either the
303  * node stops heartbeating from hb_down or the caller decides that the node is
304  * still up and calls still_up */
305 void o2quo_conn_err(u8 node)
306 {
307 	struct o2quo_state *qs = &o2quo_state;
308 
309 	spin_lock(&qs->qs_lock);
310 
311 	if (test_bit(node, qs->qs_conn_bm)) {
312 		qs->qs_connected--;
313 		mlog_bug_on_msg(qs->qs_connected < 0,
314 				"node %u, connected %d\n",
315 				node, qs->qs_connected);
316 
317 		clear_bit(node, qs->qs_conn_bm);
318 
319 		if (test_bit(node, qs->qs_hb_bm))
320 			o2quo_set_hold(qs, node);
321 	}
322 
323 	mlog(0, "node %u, %d total\n", node, qs->qs_connected);
324 
325 
326 	spin_unlock(&qs->qs_lock);
327 }
328 
329 void o2quo_init(void)
330 {
331 	struct o2quo_state *qs = &o2quo_state;
332 
333 	spin_lock_init(&qs->qs_lock);
334 	INIT_WORK(&qs->qs_work, o2quo_make_decision);
335 }
336 
337 void o2quo_exit(void)
338 {
339 	struct o2quo_state *qs = &o2quo_state;
340 
341 	flush_work(&qs->qs_work);
342 }
343