linux/net/tipc/monitor.c

834 lines
21 KiB
C
Raw Normal View History

tipc: add neighbor monitoring framework TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-06-14 08:46:22 +08:00
/*
* net/tipc/monitor.c
*
* Copyright (c) 2016, Ericsson AB
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. Neither the names of the copyright holders nor the names of its
* contributors may be used to endorse or promote products derived from
* this software without specific prior written permission.
*
* Alternatively, this software may be distributed under the terms of the
* GNU General Public License ("GPL") version 2 as published by the Free
* Software Foundation.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
* LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
#include <net/genetlink.h>
tipc: add neighbor monitoring framework TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-06-14 08:46:22 +08:00
#include "core.h"
#include "addr.h"
#include "monitor.h"
#include "bearer.h"
tipc: add neighbor monitoring framework TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-06-14 08:46:22 +08:00
#define MAX_MON_DOMAIN 64
#define MON_TIMEOUT 120000
#define MAX_PEER_DOWN_EVENTS 4
/* struct tipc_mon_domain: domain record to be transferred between peers
* @len: actual size of domain record
* @gen: current generation of sender's domain
* @ack_gen: most recent generation of self's domain acked by peer
* @member_cnt: number of domain member nodes described in this record
* @up_map: bit map indicating which of the members the sender considers up
* @members: identity of the domain members
*/
struct tipc_mon_domain {
u16 len;
u16 gen;
u16 ack_gen;
u16 member_cnt;
u64 up_map;
u32 members[MAX_MON_DOMAIN];
};
/* struct tipc_peer: state of a peer node and its domain
* @addr: tipc node identity of peer
* @head_map: shows which other nodes currently consider peer 'up'
* @domain: most recent domain record from peer
* @hash: position in hashed lookup list
* @list: position in linked list, in circular ascending order by 'addr'
* @applied: number of reported domain members applied on this monitor list
* @is_up: peer is up as seen from this node
* @is_head: peer is assigned domain head as seen from this node
* @is_local: peer is in local domain and should be continuously monitored
* @down_cnt: - numbers of other peers which have reported this on lost
*/
struct tipc_peer {
u32 addr;
struct tipc_mon_domain *domain;
struct hlist_node hash;
struct list_head list;
u8 applied;
u8 down_cnt;
bool is_up;
bool is_head;
bool is_local;
};
struct tipc_monitor {
struct hlist_head peers[NODE_HTABLE_SIZE];
int peer_cnt;
struct tipc_peer *self;
rwlock_t lock;
struct tipc_mon_domain cache;
u16 list_gen;
u16 dom_gen;
struct net *net;
struct timer_list timer;
unsigned long timer_intv;
};
static struct tipc_monitor *tipc_monitor(struct net *net, int bearer_id)
{
return tipc_net(net)->monitors[bearer_id];
}
const int tipc_max_domain_size = sizeof(struct tipc_mon_domain);
/* dom_rec_len(): actual length of domain record for transport
*/
static int dom_rec_len(struct tipc_mon_domain *dom, u16 mcnt)
{
return ((void *)&dom->members - (void *)dom) + (mcnt * sizeof(u32));
}
/* dom_size() : calculate size of own domain based on number of peers
*/
static int dom_size(int peers)
{
int i = 0;
while ((i * i) < peers)
i++;
return i < MAX_MON_DOMAIN ? i : MAX_MON_DOMAIN;
}
static void map_set(u64 *up_map, int i, unsigned int v)
{
*up_map &= ~(1ULL << i);
*up_map |= ((u64)v << i);
tipc: add neighbor monitoring framework TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-06-14 08:46:22 +08:00
}
static int map_get(u64 up_map, int i)
{
return (up_map & (1 << i)) >> i;
}
static struct tipc_peer *peer_prev(struct tipc_peer *peer)
{
return list_last_entry(&peer->list, struct tipc_peer, list);
}
static struct tipc_peer *peer_nxt(struct tipc_peer *peer)
{
return list_first_entry(&peer->list, struct tipc_peer, list);
}
static struct tipc_peer *peer_head(struct tipc_peer *peer)
{
while (!peer->is_head)
peer = peer_prev(peer);
return peer;
}
static struct tipc_peer *get_peer(struct tipc_monitor *mon, u32 addr)
{
struct tipc_peer *peer;
unsigned int thash = tipc_hashfn(addr);
hlist_for_each_entry(peer, &mon->peers[thash], hash) {
if (peer->addr == addr)
return peer;
}
return NULL;
}
static struct tipc_peer *get_self(struct net *net, int bearer_id)
{
struct tipc_monitor *mon = tipc_monitor(net, bearer_id);
return mon->self;
}
static inline bool tipc_mon_is_active(struct net *net, struct tipc_monitor *mon)
{
struct tipc_net *tn = tipc_net(net);
return mon->peer_cnt > tn->mon_threshold;
}
/* mon_identify_lost_members() : - identify amd mark potentially lost members
*/
static void mon_identify_lost_members(struct tipc_peer *peer,
struct tipc_mon_domain *dom_bef,
int applied_bef)
{
struct tipc_peer *member = peer;
struct tipc_mon_domain *dom_aft = peer->domain;
int applied_aft = peer->applied;
int i;
for (i = 0; i < applied_bef; i++) {
member = peer_nxt(member);
/* Do nothing if self or peer already see member as down */
if (!member->is_up || !map_get(dom_bef->up_map, i))
continue;
/* Loss of local node must be detected by active probing */
if (member->is_local)
continue;
/* Start probing if member was removed from applied domain */
if (!applied_aft || (applied_aft < i)) {
member->down_cnt = 1;
continue;
}
/* Member loss is confirmed if it is still in applied domain */
if (!map_get(dom_aft->up_map, i))
member->down_cnt++;
}
}
/* mon_apply_domain() : match a peer's domain record against monitor list
*/
static void mon_apply_domain(struct tipc_monitor *mon,
struct tipc_peer *peer)
{
struct tipc_mon_domain *dom = peer->domain;
struct tipc_peer *member;
u32 addr;
int i;
if (!dom || !peer->is_up)
return;
/* Scan across domain members and match against monitor list */
peer->applied = 0;
member = peer_nxt(peer);
for (i = 0; i < dom->member_cnt; i++) {
addr = dom->members[i];
if (addr != member->addr)
return;
peer->applied++;
member = peer_nxt(member);
}
}
/* mon_update_local_domain() : update after peer addition/removal/up/down
*/
static void mon_update_local_domain(struct tipc_monitor *mon)
{
struct tipc_peer *self = mon->self;
struct tipc_mon_domain *cache = &mon->cache;
struct tipc_mon_domain *dom = self->domain;
struct tipc_peer *peer = self;
u64 prev_up_map = dom->up_map;
u16 member_cnt, i;
bool diff;
/* Update local domain size based on current size of cluster */
member_cnt = dom_size(mon->peer_cnt) - 1;
self->applied = member_cnt;
/* Update native and cached outgoing local domain records */
dom->len = dom_rec_len(dom, member_cnt);
diff = dom->member_cnt != member_cnt;
dom->member_cnt = member_cnt;
for (i = 0; i < member_cnt; i++) {
peer = peer_nxt(peer);
diff |= dom->members[i] != peer->addr;
dom->members[i] = peer->addr;
map_set(&dom->up_map, i, peer->is_up);
cache->members[i] = htonl(peer->addr);
}
diff |= dom->up_map != prev_up_map;
if (!diff)
return;
dom->gen = ++mon->dom_gen;
cache->len = htons(dom->len);
cache->gen = htons(dom->gen);
cache->member_cnt = htons(member_cnt);
cache->up_map = cpu_to_be64(dom->up_map);
mon_apply_domain(mon, self);
}
/* mon_update_neighbors() : update preceding neighbors of added/removed peer
*/
static void mon_update_neighbors(struct tipc_monitor *mon,
struct tipc_peer *peer)
{
int dz, i;
dz = dom_size(mon->peer_cnt);
for (i = 0; i < dz; i++) {
mon_apply_domain(mon, peer);
peer = peer_prev(peer);
}
}
/* mon_assign_roles() : reassign peer roles after a network change
* The monitor list is consistent at this stage; i.e., each peer is monitoring
* a set of domain members as matched between domain record and the monitor list
*/
static void mon_assign_roles(struct tipc_monitor *mon, struct tipc_peer *head)
{
struct tipc_peer *peer = peer_nxt(head);
struct tipc_peer *self = mon->self;
int i = 0;
for (; peer != self; peer = peer_nxt(peer)) {
peer->is_local = false;
/* Update domain member */
if (i++ < head->applied) {
peer->is_head = false;
if (head == self)
peer->is_local = true;
continue;
}
/* Assign next domain head */
if (!peer->is_up)
continue;
if (peer->is_head)
break;
head = peer;
head->is_head = true;
i = 0;
}
mon->list_gen++;
}
void tipc_mon_remove_peer(struct net *net, u32 addr, int bearer_id)
{
struct tipc_monitor *mon = tipc_monitor(net, bearer_id);
struct tipc_peer *self = get_self(net, bearer_id);
struct tipc_peer *peer, *prev, *head;
write_lock_bh(&mon->lock);
peer = get_peer(mon, addr);
if (!peer)
goto exit;
prev = peer_prev(peer);
list_del(&peer->list);
hlist_del(&peer->hash);
kfree(peer->domain);
kfree(peer);
mon->peer_cnt--;
head = peer_head(prev);
if (head == self)
mon_update_local_domain(mon);
mon_update_neighbors(mon, prev);
/* Revert to full-mesh monitoring if we reach threshold */
if (!tipc_mon_is_active(net, mon)) {
list_for_each_entry(peer, &self->list, list) {
kfree(peer->domain);
peer->domain = NULL;
peer->applied = 0;
}
}
mon_assign_roles(mon, head);
exit:
write_unlock_bh(&mon->lock);
}
static bool tipc_mon_add_peer(struct tipc_monitor *mon, u32 addr,
struct tipc_peer **peer)
{
struct tipc_peer *self = mon->self;
struct tipc_peer *cur, *prev, *p;
p = kzalloc(sizeof(*p), GFP_ATOMIC);
*peer = p;
if (!p)
return false;
p->addr = addr;
/* Add new peer to lookup list */
INIT_LIST_HEAD(&p->list);
hlist_add_head(&p->hash, &mon->peers[tipc_hashfn(addr)]);
/* Sort new peer into iterator list, in ascending circular order */
prev = self;
list_for_each_entry(cur, &self->list, list) {
if ((addr > prev->addr) && (addr < cur->addr))
break;
if (((addr < cur->addr) || (addr > prev->addr)) &&
(prev->addr > cur->addr))
break;
prev = cur;
}
list_add_tail(&p->list, &cur->list);
mon->peer_cnt++;
mon_update_neighbors(mon, p);
return true;
}
void tipc_mon_peer_up(struct net *net, u32 addr, int bearer_id)
{
struct tipc_monitor *mon = tipc_monitor(net, bearer_id);
struct tipc_peer *self = get_self(net, bearer_id);
struct tipc_peer *peer, *head;
write_lock_bh(&mon->lock);
peer = get_peer(mon, addr);
if (!peer && !tipc_mon_add_peer(mon, addr, &peer))
goto exit;
peer->is_up = true;
head = peer_head(peer);
if (head == self)
mon_update_local_domain(mon);
mon_assign_roles(mon, head);
exit:
write_unlock_bh(&mon->lock);
}
void tipc_mon_peer_down(struct net *net, u32 addr, int bearer_id)
{
struct tipc_monitor *mon = tipc_monitor(net, bearer_id);
struct tipc_peer *self = get_self(net, bearer_id);
struct tipc_peer *peer, *head;
struct tipc_mon_domain *dom;
int applied;
write_lock_bh(&mon->lock);
peer = get_peer(mon, addr);
if (!peer) {
pr_warn("Mon: unknown link %x/%u DOWN\n", addr, bearer_id);
goto exit;
}
applied = peer->applied;
peer->applied = 0;
dom = peer->domain;
peer->domain = NULL;
if (peer->is_head)
mon_identify_lost_members(peer, dom, applied);
kfree(dom);
peer->is_up = false;
peer->is_head = false;
peer->is_local = false;
peer->down_cnt = 0;
head = peer_head(peer);
if (head == self)
mon_update_local_domain(mon);
mon_assign_roles(mon, head);
exit:
write_unlock_bh(&mon->lock);
}
/* tipc_mon_rcv - process monitor domain event message
*/
void tipc_mon_rcv(struct net *net, void *data, u16 dlen, u32 addr,
struct tipc_mon_state *state, int bearer_id)
{
struct tipc_monitor *mon = tipc_monitor(net, bearer_id);
struct tipc_mon_domain *arrv_dom = data;
struct tipc_mon_domain dom_bef;
struct tipc_mon_domain *dom;
struct tipc_peer *peer;
u16 new_member_cnt = ntohs(arrv_dom->member_cnt);
int new_dlen = dom_rec_len(arrv_dom, new_member_cnt);
u16 new_gen = ntohs(arrv_dom->gen);
u16 acked_gen = ntohs(arrv_dom->ack_gen);
bool probing = state->probing;
int i, applied_bef;
state->probing = false;
/* Sanity check received domain record */
if (dlen < dom_rec_len(arrv_dom, 0))
return;
if (dlen != dom_rec_len(arrv_dom, new_member_cnt))
return;
if ((dlen < new_dlen) || ntohs(arrv_dom->len) != new_dlen)
tipc: add neighbor monitoring framework TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-06-14 08:46:22 +08:00
return;
/* Synch generation numbers with peer if link just came up */
if (!state->synched) {
state->peer_gen = new_gen - 1;
state->acked_gen = acked_gen;
state->synched = true;
}
if (more(acked_gen, state->acked_gen))
state->acked_gen = acked_gen;
/* Drop duplicate unless we are waiting for a probe response */
if (!more(new_gen, state->peer_gen) && !probing)
return;
write_lock_bh(&mon->lock);
peer = get_peer(mon, addr);
if (!peer || !peer->is_up)
goto exit;
/* Peer is confirmed, stop any ongoing probing */
peer->down_cnt = 0;
/* Task is done for duplicate record */
if (!more(new_gen, state->peer_gen))
goto exit;
state->peer_gen = new_gen;
/* Cache current domain record for later use */
dom_bef.member_cnt = 0;
dom = peer->domain;
if (dom)
memcpy(&dom_bef, dom, dom->len);
/* Transform and store received domain record */
if (!dom || (dom->len < new_dlen)) {
kfree(dom);
dom = kmalloc(new_dlen, GFP_ATOMIC);
peer->domain = dom;
if (!dom)
goto exit;
}
dom->len = new_dlen;
dom->gen = new_gen;
dom->member_cnt = new_member_cnt;
dom->up_map = be64_to_cpu(arrv_dom->up_map);
for (i = 0; i < new_member_cnt; i++)
dom->members[i] = ntohl(arrv_dom->members[i]);
/* Update peers affected by this domain record */
applied_bef = peer->applied;
mon_apply_domain(mon, peer);
mon_identify_lost_members(peer, &dom_bef, applied_bef);
mon_assign_roles(mon, peer_head(peer));
exit:
write_unlock_bh(&mon->lock);
}
void tipc_mon_prep(struct net *net, void *data, int *dlen,
struct tipc_mon_state *state, int bearer_id)
{
struct tipc_monitor *mon = tipc_monitor(net, bearer_id);
struct tipc_mon_domain *dom = data;
u16 gen = mon->dom_gen;
u16 len;
/* Send invalid record if not active */
if (!tipc_mon_is_active(net, mon)) {
dom->len = 0;
tipc: add neighbor monitoring framework TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-06-14 08:46:22 +08:00
return;
}
tipc: add neighbor monitoring framework TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-06-14 08:46:22 +08:00
/* Send only a dummy record with ack if peer has acked our last sent */
if (likely(state->acked_gen == gen)) {
len = dom_rec_len(dom, 0);
*dlen = len;
dom->len = htons(len);
dom->gen = htons(gen);
dom->ack_gen = htons(state->peer_gen);
dom->member_cnt = 0;
return;
}
/* Send the full record */
read_lock_bh(&mon->lock);
len = ntohs(mon->cache.len);
*dlen = len;
memcpy(data, &mon->cache, len);
read_unlock_bh(&mon->lock);
dom->ack_gen = htons(state->peer_gen);
}
void tipc_mon_get_state(struct net *net, u32 addr,
struct tipc_mon_state *state,
int bearer_id)
{
struct tipc_monitor *mon = tipc_monitor(net, bearer_id);
struct tipc_peer *peer;
if (!tipc_mon_is_active(net, mon)) {
state->probing = false;
state->monitoring = true;
return;
}
tipc: add neighbor monitoring framework TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-06-14 08:46:22 +08:00
/* Used cached state if table has not changed */
if (!state->probing &&
(state->list_gen == mon->list_gen) &&
(state->acked_gen == mon->dom_gen))
return;
read_lock_bh(&mon->lock);
peer = get_peer(mon, addr);
if (peer) {
state->probing = state->acked_gen != mon->dom_gen;
state->probing |= peer->down_cnt;
state->reset |= peer->down_cnt >= MAX_PEER_DOWN_EVENTS;
state->monitoring = peer->is_local;
state->monitoring |= peer->is_head;
state->list_gen = mon->list_gen;
}
read_unlock_bh(&mon->lock);
}
static void mon_timeout(struct timer_list *t)
tipc: add neighbor monitoring framework TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-06-14 08:46:22 +08:00
{
struct tipc_monitor *mon = from_timer(mon, t, timer);
tipc: add neighbor monitoring framework TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-06-14 08:46:22 +08:00
struct tipc_peer *self;
int best_member_cnt = dom_size(mon->peer_cnt) - 1;
write_lock_bh(&mon->lock);
self = mon->self;
if (self && (best_member_cnt != self->applied)) {
mon_update_local_domain(mon);
mon_assign_roles(mon, self);
}
write_unlock_bh(&mon->lock);
mod_timer(&mon->timer, jiffies + mon->timer_intv);
}
int tipc_mon_create(struct net *net, int bearer_id)
{
struct tipc_net *tn = tipc_net(net);
struct tipc_monitor *mon;
struct tipc_peer *self;
struct tipc_mon_domain *dom;
if (tn->monitors[bearer_id])
return 0;
mon = kzalloc(sizeof(*mon), GFP_ATOMIC);
self = kzalloc(sizeof(*self), GFP_ATOMIC);
dom = kzalloc(sizeof(*dom), GFP_ATOMIC);
if (!mon || !self || !dom) {
kfree(mon);
kfree(self);
kfree(dom);
return -ENOMEM;
}
tn->monitors[bearer_id] = mon;
rwlock_init(&mon->lock);
mon->net = net;
mon->peer_cnt = 1;
mon->self = self;
self->domain = dom;
self->addr = tipc_own_addr(net);
self->is_up = true;
self->is_head = true;
INIT_LIST_HEAD(&self->list);
timer_setup(&mon->timer, mon_timeout, 0);
tipc: add neighbor monitoring framework TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-06-14 08:46:22 +08:00
mon->timer_intv = msecs_to_jiffies(MON_TIMEOUT + (tn->random & 0xffff));
mod_timer(&mon->timer, jiffies + mon->timer_intv);
return 0;
}
void tipc_mon_delete(struct net *net, int bearer_id)
{
struct tipc_net *tn = tipc_net(net);
struct tipc_monitor *mon = tipc_monitor(net, bearer_id);
tipc: fix tipc_mon_delete() oops in tipc_enable_bearer() error path Calling tipc_mon_delete() before the monitor has been created will oops. This can happen in tipc_enable_bearer() error path if tipc_disc_create() fails. [ 48.589074] BUG: unable to handle kernel paging request at 0000000000001008 [ 48.590266] IP: tipc_mon_delete+0xea/0x270 [tipc] [ 48.591223] PGD 1e60c5067 P4D 1e60c5067 PUD 1eb0cf067 PMD 0 [ 48.592230] Oops: 0000 [#1] SMP KASAN [ 48.595610] CPU: 5 PID: 1199 Comm: tipc Tainted: G B 4.15.0-rc4-pc64-dirty #5 [ 48.597176] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-2.fc27 04/01/2014 [ 48.598489] RIP: 0010:tipc_mon_delete+0xea/0x270 [tipc] [ 48.599347] RSP: 0018:ffff8801d827f668 EFLAGS: 00010282 [ 48.600705] RAX: ffff8801ee813f00 RBX: 0000000000000204 RCX: 0000000000000000 [ 48.602183] RDX: 1ffffffff1de6a75 RSI: 0000000000000297 RDI: 0000000000000297 [ 48.604373] RBP: 0000000000000000 R08: 0000000000000000 R09: fffffbfff1dd1533 [ 48.605607] R10: ffffffff8eafbb05 R11: fffffbfff1dd1534 R12: 0000000000000050 [ 48.607082] R13: dead000000000200 R14: ffffffff8e73f310 R15: 0000000000001020 [ 48.608228] FS: 00007fc686484800(0000) GS:ffff8801f5540000(0000) knlGS:0000000000000000 [ 48.610189] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 48.611459] CR2: 0000000000001008 CR3: 00000001dda70002 CR4: 00000000003606e0 [ 48.612759] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 48.613831] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 48.615038] Call Trace: [ 48.615635] tipc_enable_bearer+0x415/0x5e0 [tipc] [ 48.620623] tipc_nl_bearer_enable+0x1ab/0x200 [tipc] [ 48.625118] genl_family_rcv_msg+0x36b/0x570 [ 48.631233] genl_rcv_msg+0x5a/0xa0 [ 48.631867] netlink_rcv_skb+0x1cc/0x220 [ 48.636373] genl_rcv+0x24/0x40 [ 48.637306] netlink_unicast+0x29c/0x350 [ 48.639664] netlink_sendmsg+0x439/0x590 [ 48.642014] SYSC_sendto+0x199/0x250 [ 48.649912] do_syscall_64+0xfd/0x2c0 [ 48.650651] entry_SYSCALL64_slow_path+0x25/0x25 [ 48.651843] RIP: 0033:0x7fc6859848e3 [ 48.652539] RSP: 002b:00007ffd25dff938 EFLAGS: 00000246 ORIG_RAX: 000000000000002c [ 48.654003] RAX: ffffffffffffffda RBX: 00007ffd25dff990 RCX: 00007fc6859848e3 [ 48.655303] RDX: 0000000000000054 RSI: 00007ffd25dff990 RDI: 0000000000000003 [ 48.656512] RBP: 00007ffd25dff980 R08: 00007fc685c35fc0 R09: 000000000000000c [ 48.657697] R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000d13010 [ 48.658840] R13: 00007ffd25e009c0 R14: 0000000000000000 R15: 0000000000000000 [ 48.662972] RIP: tipc_mon_delete+0xea/0x270 [tipc] RSP: ffff8801d827f668 [ 48.664073] CR2: 0000000000001008 [ 48.664576] ---[ end trace e811818d54d5ce88 ]--- Acked-by: Ying Xue <ying.xue@windriver.com> Acked-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: Tommi Rantala <tommi.t.rantala@nokia.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-12-22 15:35:17 +08:00
struct tipc_peer *self;
tipc: add neighbor monitoring framework TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-06-14 08:46:22 +08:00
struct tipc_peer *peer, *tmp;
tipc: fix tipc_mon_delete() oops in tipc_enable_bearer() error path Calling tipc_mon_delete() before the monitor has been created will oops. This can happen in tipc_enable_bearer() error path if tipc_disc_create() fails. [ 48.589074] BUG: unable to handle kernel paging request at 0000000000001008 [ 48.590266] IP: tipc_mon_delete+0xea/0x270 [tipc] [ 48.591223] PGD 1e60c5067 P4D 1e60c5067 PUD 1eb0cf067 PMD 0 [ 48.592230] Oops: 0000 [#1] SMP KASAN [ 48.595610] CPU: 5 PID: 1199 Comm: tipc Tainted: G B 4.15.0-rc4-pc64-dirty #5 [ 48.597176] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-2.fc27 04/01/2014 [ 48.598489] RIP: 0010:tipc_mon_delete+0xea/0x270 [tipc] [ 48.599347] RSP: 0018:ffff8801d827f668 EFLAGS: 00010282 [ 48.600705] RAX: ffff8801ee813f00 RBX: 0000000000000204 RCX: 0000000000000000 [ 48.602183] RDX: 1ffffffff1de6a75 RSI: 0000000000000297 RDI: 0000000000000297 [ 48.604373] RBP: 0000000000000000 R08: 0000000000000000 R09: fffffbfff1dd1533 [ 48.605607] R10: ffffffff8eafbb05 R11: fffffbfff1dd1534 R12: 0000000000000050 [ 48.607082] R13: dead000000000200 R14: ffffffff8e73f310 R15: 0000000000001020 [ 48.608228] FS: 00007fc686484800(0000) GS:ffff8801f5540000(0000) knlGS:0000000000000000 [ 48.610189] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 48.611459] CR2: 0000000000001008 CR3: 00000001dda70002 CR4: 00000000003606e0 [ 48.612759] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 48.613831] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 48.615038] Call Trace: [ 48.615635] tipc_enable_bearer+0x415/0x5e0 [tipc] [ 48.620623] tipc_nl_bearer_enable+0x1ab/0x200 [tipc] [ 48.625118] genl_family_rcv_msg+0x36b/0x570 [ 48.631233] genl_rcv_msg+0x5a/0xa0 [ 48.631867] netlink_rcv_skb+0x1cc/0x220 [ 48.636373] genl_rcv+0x24/0x40 [ 48.637306] netlink_unicast+0x29c/0x350 [ 48.639664] netlink_sendmsg+0x439/0x590 [ 48.642014] SYSC_sendto+0x199/0x250 [ 48.649912] do_syscall_64+0xfd/0x2c0 [ 48.650651] entry_SYSCALL64_slow_path+0x25/0x25 [ 48.651843] RIP: 0033:0x7fc6859848e3 [ 48.652539] RSP: 002b:00007ffd25dff938 EFLAGS: 00000246 ORIG_RAX: 000000000000002c [ 48.654003] RAX: ffffffffffffffda RBX: 00007ffd25dff990 RCX: 00007fc6859848e3 [ 48.655303] RDX: 0000000000000054 RSI: 00007ffd25dff990 RDI: 0000000000000003 [ 48.656512] RBP: 00007ffd25dff980 R08: 00007fc685c35fc0 R09: 000000000000000c [ 48.657697] R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000d13010 [ 48.658840] R13: 00007ffd25e009c0 R14: 0000000000000000 R15: 0000000000000000 [ 48.662972] RIP: tipc_mon_delete+0xea/0x270 [tipc] RSP: ffff8801d827f668 [ 48.664073] CR2: 0000000000001008 [ 48.664576] ---[ end trace e811818d54d5ce88 ]--- Acked-by: Ying Xue <ying.xue@windriver.com> Acked-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: Tommi Rantala <tommi.t.rantala@nokia.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2017-12-22 15:35:17 +08:00
if (!mon)
return;
self = get_self(net, bearer_id);
tipc: add neighbor monitoring framework TIPC based clusters are by default set up with full-mesh link connectivity between all nodes. Those links are expected to provide a short failure detection time, by default set to 1500 ms. Because of this, the background load for neighbor monitoring in an N-node cluster increases with a factor N on each node, while the overall monitoring traffic through the network infrastructure increases at a ~(N * (N - 1)) rate. Experience has shown that such clusters don't scale well beyond ~100 nodes unless we significantly increase failure discovery tolerance. This commit introduces a framework and an algorithm that drastically reduces this background load, while basically maintaining the original failure detection times across the whole cluster. Using this algorithm, background load will now grow at a rate of ~(2 * sqrt(N)) per node, and at ~(2 * N * sqrt(N)) in traffic overhead. As an example, each node will now have to actively monitor 38 neighbors in a 400-node cluster, instead of as before 399. This "Overlapping Ring Supervision Algorithm" is completely distributed and employs no centralized or coordinated state. It goes as follows: - Each node makes up a linearly ascending, circular list of all its N known neighbors, based on their TIPC node identity. This algorithm must be the same on all nodes. - The node then selects the next M = sqrt(N) - 1 nodes downstream from itself in the list, and chooses to actively monitor those. This is called its "local monitoring domain". - It creates a domain record describing the monitoring domain, and piggy-backs this in the data area of all neighbor monitoring messages (LINK_PROTOCOL/STATE) leaving that node. This means that all nodes in the cluster eventually (default within 400 ms) will learn about its monitoring domain. - Whenever a node discovers a change in its local domain, e.g., a node has been added or has gone down, it creates and sends out a new version of its node record to inform all neighbors about the change. - A node receiving a domain record from anybody outside its local domain matches this against its own list (which may not look the same), and chooses to not actively monitor those members of the received domain record that are also present in its own list. Instead, it relies on indications from the direct monitoring nodes if an indirectly monitored node has gone up or down. If a node is indicated lost, the receiving node temporarily activates its own direct monitoring towards that node in order to confirm, or not, that it is actually gone. - Since each node is actively monitoring sqrt(N) downstream neighbors, each node is also actively monitored by the same number of upstream neighbors. This means that all non-direct monitoring nodes normally will receive sqrt(N) indications that a node is gone. - A major drawback with ring monitoring is how it handles failures that cause massive network partitionings. If both a lost node and all its direct monitoring neighbors are inside the lost partition, the nodes in the remaining partition will never receive indications about the loss. To overcome this, each node also chooses to actively monitor some nodes outside its local domain. Those nodes are called remote domain "heads", and are selected in such a way that no node in the cluster will be more than two direct monitoring hops away. Because of this, each node, apart from monitoring the member of its local domain, will also typically monitor sqrt(N) remote head nodes. - As an optimization, local list status, domain status and domain records are marked with a generation number. This saves senders from unnecessarily conveying unaltered domain records, and receivers from performing unneeded re-adaptations of their node monitoring list, such as re-assigning domain heads. - As a measure of caution we have added the possibility to disable the new algorithm through configuration. We do this by keeping a threshold value for the cluster size; a cluster that grows beyond this value will switch from full-mesh to ring monitoring, and vice versa when it shrinks below the value. This means that if the threshold is set to a value larger than any anticipated cluster size (default size is 32) the new algorithm is effectively disabled. A patch set for altering the threshold value and for listing the table contents will follow shortly. - This change is fully backwards compatible. Acked-by: Ying Xue <ying.xue@windriver.com> Signed-off-by: Jon Maloy <jon.maloy@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2016-06-14 08:46:22 +08:00
write_lock_bh(&mon->lock);
tn->monitors[bearer_id] = NULL;
list_for_each_entry_safe(peer, tmp, &self->list, list) {
list_del(&peer->list);
hlist_del(&peer->hash);
kfree(peer->domain);
kfree(peer);
}
mon->self = NULL;
write_unlock_bh(&mon->lock);
del_timer_sync(&mon->timer);
kfree(self->domain);
kfree(self);
kfree(mon);
}
void tipc_mon_reinit_self(struct net *net)
{
struct tipc_monitor *mon;
int bearer_id;
for (bearer_id = 0; bearer_id < MAX_BEARERS; bearer_id++) {
mon = tipc_monitor(net, bearer_id);
if (!mon)
continue;
write_lock_bh(&mon->lock);
mon->self->addr = tipc_own_addr(net);
write_unlock_bh(&mon->lock);
}
}
int tipc_nl_monitor_set_threshold(struct net *net, u32 cluster_size)
{
struct tipc_net *tn = tipc_net(net);
if (cluster_size > TIPC_CLUSTER_SIZE)
return -EINVAL;
tn->mon_threshold = cluster_size;
return 0;
}
int tipc_nl_monitor_get_threshold(struct net *net)
{
struct tipc_net *tn = tipc_net(net);
return tn->mon_threshold;
}
static int __tipc_nl_add_monitor_peer(struct tipc_peer *peer,
struct tipc_nl_msg *msg)
{
struct tipc_mon_domain *dom = peer->domain;
struct nlattr *attrs;
void *hdr;
hdr = genlmsg_put(msg->skb, msg->portid, msg->seq, &tipc_genl_family,
NLM_F_MULTI, TIPC_NL_MON_PEER_GET);
if (!hdr)
return -EMSGSIZE;
attrs = nla_nest_start_noflag(msg->skb, TIPC_NLA_MON_PEER);
if (!attrs)
goto msg_full;
if (nla_put_u32(msg->skb, TIPC_NLA_MON_PEER_ADDR, peer->addr))
goto attr_msg_full;
if (nla_put_u32(msg->skb, TIPC_NLA_MON_PEER_APPLIED, peer->applied))
goto attr_msg_full;
if (peer->is_up)
if (nla_put_flag(msg->skb, TIPC_NLA_MON_PEER_UP))
goto attr_msg_full;
if (peer->is_local)
if (nla_put_flag(msg->skb, TIPC_NLA_MON_PEER_LOCAL))
goto attr_msg_full;
if (peer->is_head)
if (nla_put_flag(msg->skb, TIPC_NLA_MON_PEER_HEAD))
goto attr_msg_full;
if (dom) {
if (nla_put_u32(msg->skb, TIPC_NLA_MON_PEER_DOMGEN, dom->gen))
goto attr_msg_full;
if (nla_put_u64_64bit(msg->skb, TIPC_NLA_MON_PEER_UPMAP,
dom->up_map, TIPC_NLA_MON_PEER_PAD))
goto attr_msg_full;
if (nla_put(msg->skb, TIPC_NLA_MON_PEER_MEMBERS,
dom->member_cnt * sizeof(u32), &dom->members))
goto attr_msg_full;
}
nla_nest_end(msg->skb, attrs);
genlmsg_end(msg->skb, hdr);
return 0;
attr_msg_full:
nla_nest_cancel(msg->skb, attrs);
msg_full:
genlmsg_cancel(msg->skb, hdr);
return -EMSGSIZE;
}
int tipc_nl_add_monitor_peer(struct net *net, struct tipc_nl_msg *msg,
u32 bearer_id, u32 *prev_node)
{
struct tipc_monitor *mon = tipc_monitor(net, bearer_id);
struct tipc_peer *peer;
if (!mon)
return -EINVAL;
read_lock_bh(&mon->lock);
peer = mon->self;
do {
if (*prev_node) {
if (peer->addr == *prev_node)
*prev_node = 0;
else
continue;
}
if (__tipc_nl_add_monitor_peer(peer, msg)) {
*prev_node = peer->addr;
read_unlock_bh(&mon->lock);
return -EMSGSIZE;
}
} while ((peer = peer_nxt(peer)) != mon->self);
read_unlock_bh(&mon->lock);
return 0;
}
int __tipc_nl_add_monitor(struct net *net, struct tipc_nl_msg *msg,
u32 bearer_id)
{
struct tipc_monitor *mon = tipc_monitor(net, bearer_id);
char bearer_name[TIPC_MAX_BEARER_NAME];
struct nlattr *attrs;
void *hdr;
int ret;
ret = tipc_bearer_get_name(net, bearer_name, bearer_id);
if (ret || !mon)
return 0;
hdr = genlmsg_put(msg->skb, msg->portid, msg->seq, &tipc_genl_family,
NLM_F_MULTI, TIPC_NL_MON_GET);
if (!hdr)
return -EMSGSIZE;
attrs = nla_nest_start_noflag(msg->skb, TIPC_NLA_MON);
if (!attrs)
goto msg_full;
read_lock_bh(&mon->lock);
if (nla_put_u32(msg->skb, TIPC_NLA_MON_REF, bearer_id))
goto attr_msg_full;
if (tipc_mon_is_active(net, mon))
if (nla_put_flag(msg->skb, TIPC_NLA_MON_ACTIVE))
goto attr_msg_full;
if (nla_put_string(msg->skb, TIPC_NLA_MON_BEARER_NAME, bearer_name))
goto attr_msg_full;
if (nla_put_u32(msg->skb, TIPC_NLA_MON_PEERCNT, mon->peer_cnt))
goto attr_msg_full;
if (nla_put_u32(msg->skb, TIPC_NLA_MON_LISTGEN, mon->list_gen))
goto attr_msg_full;
read_unlock_bh(&mon->lock);
nla_nest_end(msg->skb, attrs);
genlmsg_end(msg->skb, hdr);
return 0;
attr_msg_full:
read_unlock_bh(&mon->lock);
nla_nest_cancel(msg->skb, attrs);
msg_full:
genlmsg_cancel(msg->skb, hdr);
return -EMSGSIZE;
}