hb_block->hb_seq = cpu_to_le64(cputime);
        hb_block->hb_node = node_num;
        hb_block->hb_generation = cpu_to_le64(generation);
+       hb_block->hb_dead_ms = cpu_to_le32(o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS);
 
        /* This step must always happen last! */
        hb_block->hb_cksum = cpu_to_le32(o2hb_compute_block_crc_le(reg,
        struct o2nm_node *node;
        struct o2hb_disk_heartbeat_block *hb_block = reg->hr_tmp_block;
        u64 cputime;
+       unsigned int dead_ms = o2hb_dead_threshold * O2HB_REGION_TIMEOUT_MS;
+       unsigned int slot_dead_ms;
 
        memcpy(hb_block, slot->ds_raw_block, reg->hr_block_bytes);
 
                              &o2hb_live_slots[slot->ds_node_num]);
 
                slot->ds_equal_samples = 0;
+
+               /* We want to be sure that all nodes agree on the
+                * number of milliseconds before a node will be
+                * considered dead. The self-fencing timeout is
+                * computed from this value, and a discrepancy might
+                * result in heartbeat calling a node dead when it
+                * hasn't self-fenced yet. */
+               slot_dead_ms = le32_to_cpu(hb_block->hb_dead_ms);
+               if (slot_dead_ms && slot_dead_ms != dead_ms) {
+                       /* TODO: Perhaps we can fail the region here. */
+                       mlog(ML_ERROR, "Node %d on device %s has a dead count "
+                            "of %u ms, but our count is %u ms.\n"
+                            "Please double check your configuration values "
+                            "for 'O2CB_HEARTBEAT_THRESHOLD'\n",
+                            slot->ds_node_num, reg->hr_dev_name, slot_dead_ms,
+                            dead_ms);
+               }
                goto out;
        }