View Source global (kernel v10.0)
A global name registration facility.
This module consists of the following services:
- Registration of global names
- Global locks
- Maintenance of the fully connected network
As of OTP 25, global
will by default prevent overlapping partitions due to
network issues by actively disconnecting from nodes that reports that they have
lost connections to other nodes. This will cause fully connected partitions to
form instead of leaving the network in a state with overlapping partitions.
Warning
Prevention of overlapping partitions can be disabled using the
prevent_overlapping_partitions
Kernel parameter, makingglobal
behave like it used to do. This is, however, problematic for all applications expecting a fully connected network to be provided, such as for examplemnesia
, but also forglobal
itself. A network of overlapping partitions might cause the internal state ofglobal
to become inconsistent. Such an inconsistency can remain even after such partitions have been brought together to form a fully connected network again. The effect on other applications that expects that a fully connected network is maintained may vary, but they might misbehave in very subtle hard to detect ways during such a partitioning. Since you might get hard to detect issues without this fix, you are strongly advised not to disable this fix. Also note that this fix has to be enabled on all nodes in the network in order to work properly.
Note
None of the above services will be reliably delivered unless both of the kernel parameters
connect_all
andprevent_overlapping_partitions
are enabled. Calls to theglobal
API will, however, not fail even though one or both of them are disabled. You will just get unreliable results.
These services are controlled through the process global_name_server
that
exists on every node. The global name server starts automatically when a node is
started. With the term global is meant over a system consisting of many Erlang
nodes.
The ability to globally register names is a central concept in the programming
of distributed Erlang systems. In this module, the equivalent of the
register/2
and whereis/1
BIFs (for local name
registration) are provided, but for a network of Erlang nodes. A registered name
is an alias for a process identifier (pid). The global name server monitors
globally registered pids. If a process terminates, the name is also globally
unregistered.
The registered names are stored in replica global name tables on every node. There is no central storage point. Thus, the translation of a name to a pid is fast, as it is always done locally. For any action resulting in a change to the global name table, all tables on other nodes are automatically updated.
Global locks have lock identities and are set on a specific resource. For example, the specified resource can be a pid. When a global lock is set, access to the locked resource is denied for all resources other than the lock requester.
Both the registration and lock services are atomic. All nodes involved in these actions have the same view of the information.
The global name server also performs the critical task of continuously
monitoring changes in node configuration. If a node that runs a globally
registered process goes down, the name is globally unregistered. To this end,
the global name server subscribes to nodeup
and nodedown
messages sent from
module net_kernel
. Relevant Kernel application variables in this context are
net_setuptime
, net_ticktime
,
and dist_auto_connect
.
The name server also maintains a fully connected network. For example, if node
N1
connects to node N2
(which is already connected to N3
), the global name
servers on the nodes N1
and N3
ensure that also N1
and N3
are connected.
In this case, the name registration service cannot be used, but the lock
mechanism still works.
If the global name server fails to connect nodes (N1
and N3
in the example),
a warning event is sent to the error logger. The presence of such an event does
not exclude the nodes to connect later (you can, for example, try command
rpc:call(N1, net_adm, ping, [N2])
in the Erlang shell), but it indicates a
network problem.
Note
If the fully connected network is not set up properly, try first to increase the value of
net_setuptime
.
See Also
Summary
Types
A lock id used to set or delete lock ResourceId
on behalf of LockRequesterId
.
Functions
Equivalent to del_lock(Id, [node() | nodes()])
.
Deletes the lock Id
synchronously.
Disconnect from all other nodes known to global
.
The function unregisters both pids and sends the message
{global_name_conflict, Name, OtherPid}
to both processes.
The function randomly selects one of the pids for registration and kills the other one.
The function randomly selects one of the pids for registration, and sends the
message {global_name_conflict, Name}
to the other pid.
Atomically changes the registered name Name
on all nodes to refer to Pid
.
Globally associates name Name
with a pid, that is, globally notifies all nodes
of a new global name in a network of Erlang nodes.
Returns a list of all globally registered names.
Sends message Msg
to the pid globally registered as Name
.
Equivalent to set_lock(Id, [node() | nodes()], infinity)
.
Equivalent to set_lock(Id, Nodes, infinity)
.
Sets a lock on the specified nodes on using id/0
.
Synchronizes the global name server with all nodes known to this node.
Equivalent to trans(Id, Fun, [node() | nodes()], infinity)
.
Equivalent to trans(Id, Fun, Nodes, infinity)
.
Sets a lock on Id
(using set_lock/3
).
Removes the globally registered name Name
from the network of Erlang nodes.
Returns the pid with the globally registered name Name
. Returns undefined
if
the name is not globally registered.
Types
A lock id used to set or delete lock ResourceId
on behalf of LockRequesterId
.
-type retries() :: non_neg_integer() | infinity.
Functions
-spec del_lock(Id) -> true when Id :: id().
Equivalent to del_lock(Id, [node() | nodes()])
.
Deletes the lock Id
synchronously.
-spec disconnect() -> [node()].
Disconnect from all other nodes known to global
.
A list of node names (in an unspecified order) is returned which corresponds to
the nodes that were disconnected. All disconnect operations performed have completed when
global:disconnect/0
returns.
The disconnects will be made in such a way that only the current node will be
removed from the cluster of global
nodes. If
prevent_overlapping_partitions
is
enabled and you disconnect, from other nodes in the cluster of global
nodes,
by other means, global
on the other nodes may partition the remaining nodes in
order to ensure that no overlapping partitions appear. Even if
prevent_overlapping_partitions
is disabled, you should preferably use
global:disconnect/0
in order to remove current node from a cluster of global
nodes, since you otherwise likely will create overlapping partitions which
might cause problems.
Note that if the node is going to be halted, there is no need to remove it
from a cluster of global
nodes explicitly by calling global:disconnect/0
before halting it. The removal from the cluster is taken care of automatically
when the node halts regardless of whether prevent_overlapping_partitions
is
enabled or not.
If current node has been configured to be part of a
global group, only connected and/or synchronized nodes in
that group are known to global
, so global:disconnect/0
will only
disconnect from those nodes. If current node is not part of a global group,
all connected visible nodes will be known to global
, so
global:disconnect/0
will disconnect from all those nodes.
Note that information about connected nodes does not instantaneously reach
global
, so the caller might see a node part of the result returned by
nodes()
while it still is not known to global
. The
disconnect operation will, however, still not cause any overlapping partitions
when prevent_overlapping_partitions
is enabled. If
prevent_overlapping_partitions
is disabled, overlapping partitions might form
in this case.
Note that when prevent_overlapping_partitions
is enabled, you may see warning
reports on other nodes when they detect that current node has disconnected.
These are in this case completely harmless and can be ignored.
The function unregisters both pids and sends the message
{global_name_conflict, Name, OtherPid}
to both processes.
Can be used as a name resolving function for register_name/3
and
re_register_name/3
.
-spec random_exit_name(Name, Pid1, Pid2) -> pid() when Name :: term(), Pid1 :: pid(), Pid2 :: pid().
The function randomly selects one of the pids for registration and kills the other one.
Can be used as a name resolving function for register_name/3
and
re_register_name/3
.
-spec random_notify_name(Name, Pid1, Pid2) -> pid() when Name :: term(), Pid1 :: pid(), Pid2 :: pid().
The function randomly selects one of the pids for registration, and sends the
message {global_name_conflict, Name}
to the other pid.
Can be used as a name resolving function for register_name/3
and
re_register_name/3
.
Equivalent to re_register_name(Name, Pid, fun random_exit_name/3)
.
-spec re_register_name(Name, Pid, Resolve) -> yes when Name :: term(), Pid :: pid(), Resolve :: method().
Atomically changes the registered name Name
on all nodes to refer to Pid
.
Function Resolve
has the same behavior as in
register_name/2,3
.
Equivalent to register_name(Name, Pid, fun random_exit_name/3)
.
-spec register_name(Name, Pid, Resolve) -> yes | no when Name :: term(), Pid :: pid(), Resolve :: method().
Globally associates name Name
with a pid, that is, globally notifies all nodes
of a new global name in a network of Erlang nodes.
When new nodes are added to the network, they are informed of the globally
registered names that already exist. The network is also informed of any global
names in newly connected nodes. If any name clashes are discovered, function
Resolve
is called. Its purpose is to decide which pid is correct. If the
function crashes, or returns anything other than one of the pids, the name is
unregistered. This function is called once for each name clash.
Warning
If you plan to change code without restarting your system, you must use an external fun (
fun Module:Function/Arity
) as functionResolve
. If you use a local fun, you can never replace the code for the module that the fun belongs to.
Three predefined resolve functions exist:
random_exit_name/3
,
random_notify_name/3
, and
notify_all_name/3
.
This function is completely synchronous, that is, when this function returns, the name is either registered on all nodes or none.
The function returns yes
if successful, no
if it fails. For example, no
is
returned if an attempt is made to register an already registered process or to
register a process with a name that is already in use.
Note
Releases up to and including Erlang/OTP R10 did not check if the process was already registered. The global name table could therefore become inconsistent. The old (buggy) behavior can be chosen by giving the Kernel application variable
global_multi_name_action
the valueallow
.
If a process with a registered name dies, or the node goes down, the name is unregistered on all nodes.
-spec registered_names() -> [Name] when Name :: term().
Returns a list of all globally registered names.
Sends message Msg
to the pid globally registered as Name
.
If Name
is not a globally registered name, the calling function exits with
reason {badarg, {Name, Msg}}
.
Equivalent to set_lock(Id, [node() | nodes()], infinity)
.
Equivalent to set_lock(Id, Nodes, infinity)
.
-spec set_lock(Id, Nodes, Retries) -> boolean() when Id :: id(), Nodes :: [node()], Retries :: retries().
Sets a lock on the specified nodes on using id/0
.
If a lock already exists on ResourceId
for another requester than LockRequesterId
,
and Retries
is not equal to 0
, the process sleeps for a while and tries to
execute the action later. When Retries
attempts have been made, false
is
returned, otherwise true
. If Retries
is infinity
, true
is eventually
returned (unless the lock is never released).
This function is completely synchronous.
If a process that holds a lock dies, or the node goes down, the locks held by the process are deleted.
The global name server keeps track of all processes sharing the same lock, that is, if two processes set the same lock, both processes must delete the lock.
This function does not address the problem of a deadlock. A deadlock can never occur as long as processes only lock one resource at a time. A deadlock can occur if some processes try to lock two or more resources. It is up to the application to detect and rectify a deadlock.
Note
Avoid the following values of
ResourceId
, otherwise Erlang/OTP does not work properly:
dist_ac
global
mnesia_adjust_log_writes
mnesia_table_lock
-spec sync() -> ok | {error, Reason :: term()}.
Synchronizes the global name server with all nodes known to this node.
These are the nodes that are returned from nodes()
. When
this function returns, the global name server receives global information from
all nodes. This function can be called when new nodes are added to the network.
The only possible error reason Reason
is
{"global_groups definition error", Error}
.
Equivalent to trans(Id, Fun, [node() | nodes()], infinity)
.
-spec trans(Id, Fun, Nodes) -> Res | aborted when Id :: id(), Fun :: trans_fun(), Nodes :: [node()], Res :: term().
Equivalent to trans(Id, Fun, Nodes, infinity)
.
-spec trans(Id, Fun, Nodes, Retries) -> Res | aborted when Id :: id(), Fun :: trans_fun(), Nodes :: [node()], Retries :: retries(), Res :: term().
Sets a lock on Id
(using set_lock/3
).
If this succeeds, Fun()
is evaluated and the result Res
is returned.
Returns aborted
if the lock attempt fails. If Retries
is set to infinity
,
the transaction does not abort.
infinity
is the default setting and is used if no value is specified for
Retries
.
-spec unregister_name(Name) -> _ when Name :: term().
Removes the globally registered name Name
from the network of Erlang nodes.
Returns the pid with the globally registered name Name
. Returns undefined
if
the name is not globally registered.