fs/ocfs2/cluster/quorum.c

   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
  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         panic("ocfs2 is very sorry to be fencing this system by panicing\n");
  76 }
  77
  78 /* Indicate that a timeout occured on a hearbeat region write. The
  79  * other nodes in the cluster may consider us dead at that time so we
  80  * want to "fence" ourselves so that we don't scribble on the disk
  81  * after they think they've recovered us. This can't solve all
  82  * problems related to writeout after recovery but this hack can at
  83  * least close some of those gaps. When we have real fencing, this can
  84  * go away as our node would be fenced externally before other nodes
  85  * begin recovery. */
  86 void o2quo_disk_timeout(void)
  87 {
  88         o2quo_fence_self();
  89 }
  90
  91 static void o2quo_make_decision(void *arg)
  92 {
  93         int quorum;
  94         int lowest_hb, lowest_reachable = 0, fence = 0;
  95         struct o2quo_state *qs = &o2quo_state;
  96
  97         spin_lock(&qs->qs_lock);
  98
  99         lowest_hb = find_first_bit(qs->qs_hb_bm, O2NM_MAX_NODES);
 100         if (lowest_hb != O2NM_MAX_NODES)
 101                 lowest_reachable = test_bit(lowest_hb, qs->qs_conn_bm);
 102
 103         mlog(0, "heartbeating: %d, connected: %d, "
 104              "lowest: %d (%sreachable)\n", qs->qs_heartbeating,
 105              qs->qs_connected, lowest_hb, lowest_reachable ? "" : "un");
 106
 107         if (!test_bit(o2nm_this_node(), qs->qs_hb_bm) ||
 108             qs->qs_heartbeating == 1)
 109                 goto out;
 110
 111         if (qs->qs_heartbeating & 1) {
 112                 /* the odd numbered cluster case is straight forward --
 113                  * if we can't talk to the majority we're hosed */
 114                 quorum = (qs->qs_heartbeating + 1)/2;
 115                 if (qs->qs_connected < quorum) {
 116                         mlog(ML_ERROR, "fencing this node because it is "
 117                              "only connected to %u nodes and %u is needed "
 118                              "to make a quorum out of %u heartbeating nodes\n",
 119                              qs->qs_connected, quorum,
 120                              qs->qs_heartbeating);
 121                         fence = 1;
 122                 }
 123         } else {
 124                 /* the even numbered cluster adds the possibility of each half
 125                  * of the cluster being able to talk amongst themselves.. in
 126                  * that case we're hosed if we can't talk to the group that has
 127                  * the lowest numbered node */
 128                 quorum = qs->qs_heartbeating / 2;
 129                 if (qs->qs_connected < quorum) {
 130                         mlog(ML_ERROR, "fencing this node because it is "
 131                              "only connected to %u nodes and %u is needed "
 132                              "to make a quorum out of %u heartbeating nodes\n",
 133                              qs->qs_connected, quorum,
 134                              qs->qs_heartbeating);
 135                         fence = 1;
 136                 }
 137                 else if ((qs->qs_connected == quorum) &&
 138                          !lowest_reachable) {
 139                         mlog(ML_ERROR, "fencing this node because it is "
 140                              "connected to a half-quorum of %u out of %u "
 141                              "nodes which doesn't include the lowest active "
 142                              "node %u\n", quorum, qs->qs_heartbeating,
 143                              lowest_hb);
 144                         fence = 1;
 145                 }
 146         }
 147
 148 out:
 149         spin_unlock(&qs->qs_lock);
 150         if (fence)
 151                 o2quo_fence_self();
 152 }
 153
 154 static void o2quo_set_hold(struct o2quo_state *qs, u8 node)
 155 {
 156         assert_spin_locked(&qs->qs_lock);
 157
 158         if (!test_and_set_bit(node, qs->qs_hold_bm)) {
 159                 qs->qs_holds++;
 160                 mlog_bug_on_msg(qs->qs_holds == O2NM_MAX_NODES,
 161                                 "node %u\n", node);
 162                 mlog(0, "node %u, %d total\n", node, qs->qs_holds);
 163         }
 164 }
 165
 166 static void o2quo_clear_hold(struct o2quo_state *qs, u8 node)
 167 {
 168         assert_spin_locked(&qs->qs_lock);
 169
 170         if (test_and_clear_bit(node, qs->qs_hold_bm)) {
 171                 mlog(0, "node %u, %d total\n", node, qs->qs_holds - 1);
 172                 if (--qs->qs_holds == 0) {
 173                         if (qs->qs_pending) {
 174                                 qs->qs_pending = 0;
 175                                 schedule_work(&qs->qs_work);
 176                         }
 177                 }
 178                 mlog_bug_on_msg(qs->qs_holds < 0, "node %u, holds %d\n",
 179                                 node, qs->qs_holds);
 180         }
 181 }
 182
 183 /* as a node comes up we delay the quorum decision until we know the fate of
 184  * the connection.  the hold will be droped in conn_up or hb_down.  it might be
 185  * perpetuated by con_err until hb_down.  if we already have a conn, we might
 186  * be dropping a hold that conn_up got. */
 187 void o2quo_hb_up(u8 node)
 188 {
 189         struct o2quo_state *qs = &o2quo_state;
 190
 191         spin_lock(&qs->qs_lock);
 192
 193         qs->qs_heartbeating++;
 194         mlog_bug_on_msg(qs->qs_heartbeating == O2NM_MAX_NODES,
 195                         "node %u\n", node);
 196         mlog_bug_on_msg(test_bit(node, qs->qs_hb_bm), "node %u\n", node);
 197         set_bit(node, qs->qs_hb_bm);
 198
 199         mlog(0, "node %u, %d total\n", node, qs->qs_heartbeating);
 200
 201         if (!test_bit(node, qs->qs_conn_bm))
 202                 o2quo_set_hold(qs, node);
 203         else
 204                 o2quo_clear_hold(qs, node);
 205
 206         spin_unlock(&qs->qs_lock);
 207 }
 208
 209 /* hb going down releases any holds we might have had due to this node from
 210  * conn_up, conn_err, or hb_up */
 211 void o2quo_hb_down(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 < 0,
 219                         "node %u, %d heartbeating\n",
 220                         node, qs->qs_heartbeating);
 221         mlog_bug_on_msg(!test_bit(node, qs->qs_hb_bm), "node %u\n", node);
 222         clear_bit(node, qs->qs_hb_bm);
 223
 224         mlog(0, "node %u, %d total\n", node, qs->qs_heartbeating);
 225
 226         o2quo_clear_hold(qs, node);
 227
 228         spin_unlock(&qs->qs_lock);
 229 }
 230
 231 /* this tells us that we've decided that the node is still heartbeating
 232  * even though we've lost it's conn.  it must only be called after conn_err
 233  * and indicates that we must now make a quorum decision in the future,
 234  * though we might be doing so after waiting for holds to drain.  Here
 235  * we'll be dropping the hold from conn_err. */
 236 void o2quo_hb_still_up(u8 node)
 237 {
 238         struct o2quo_state *qs = &o2quo_state;
 239
 240         spin_lock(&qs->qs_lock);
 241
 242         mlog(0, "node %u\n", node);
 243
 244         qs->qs_pending = 1;
 245         o2quo_clear_hold(qs, node);
 246
 247         spin_unlock(&qs->qs_lock);
 248 }
 249
 250 /* This is analagous to hb_up.  as a node's connection comes up we delay the
 251  * quorum decision until we see it heartbeating.  the hold will be droped in
 252  * hb_up or hb_down.  it might be perpetuated by con_err until hb_down.  if
 253  * it's already heartbeating we we might be dropping a hold that conn_up got.
 254  * */
 255 void o2quo_conn_up(u8 node)
 256 {
 257         struct o2quo_state *qs = &o2quo_state;
 258
 259         spin_lock(&qs->qs_lock);
 260
 261         qs->qs_connected++;
 262         mlog_bug_on_msg(qs->qs_connected == O2NM_MAX_NODES,
 263                         "node %u\n", node);
 264         mlog_bug_on_msg(test_bit(node, qs->qs_conn_bm), "node %u\n", node);
 265         set_bit(node, qs->qs_conn_bm);
 266
 267         mlog(0, "node %u, %d total\n", node, qs->qs_connected);
 268
 269         if (!test_bit(node, qs->qs_hb_bm))
 270                 o2quo_set_hold(qs, node);
 271         else
 272                 o2quo_clear_hold(qs, node);
 273
 274         spin_unlock(&qs->qs_lock);
 275 }
 276
 277 /* we've decided that we won't ever be connecting to the node again.  if it's
 278  * still heartbeating we grab a hold that will delay decisions until either the
 279  * node stops heartbeating from hb_down or the caller decides that the node is
 280  * still up and calls still_up */
 281 void o2quo_conn_err(u8 node)
 282 {
 283         struct o2quo_state *qs = &o2quo_state;
 284
 285         spin_lock(&qs->qs_lock);
 286
 287         if (test_bit(node, qs->qs_conn_bm)) {
 288                 qs->qs_connected--;
 289                 mlog_bug_on_msg(qs->qs_connected < 0,
 290                                 "node %u, connected %d\n",
 291                                 node, qs->qs_connected);
 292
 293                 clear_bit(node, qs->qs_conn_bm);
 294         }
 295
 296         mlog(0, "node %u, %d total\n", node, qs->qs_connected);
 297
 298         if (test_bit(node, qs->qs_hb_bm))
 299                 o2quo_set_hold(qs, node);
 300
 301         spin_unlock(&qs->qs_lock);
 302 }
 303
 304 void o2quo_init(void)
 305 {
 306         struct o2quo_state *qs = &o2quo_state;
 307
 308         spin_lock_init(&qs->qs_lock);
 309         INIT_WORK(&qs->qs_work, o2quo_make_decision, NULL);
 310 }
 311
 312 void o2quo_exit(void)
 313 {
 314         flush_scheduled_work();
 315 }