icinga2/doc/19-technical-concepts.md

Technical Concepts

This chapter provides technical concepts and design insights into specific Icinga 2 components such as:

• Application
• Configuration
• Features
• Check Scheduler
• Cluster
• TLS Network IO

Application

CLI Commands

The Icinga 2 application is managed with different CLI sub commands. daemon takes care about loading the configuration files, running the application as daemon, etc. Other sub commands allow to enable features, generate and request TLS certificates or enter the debug console.

The main entry point for each CLI command parses the command line parameters and then triggers the required actions.

daemon CLI command

This CLI command loads the configuration files, starting with icinga2.conf. The configuration compiler parses the file and detects additional file includes, constants, and any other DSL specific declaration.

At this stage, the configuration will already be checked against the defined grammar in the scanner, and custom object validators will also be checked.

If the user provided -C/--validate, the CLI command returns with the validation exit code.

When running as daemon, additional parameters are checked, e.g. whether this application was triggered by a reload, needs to daemonize with fork() involved and update the object's authority. The latter is important for HA-enabled cluster zones.

Configuration

Lexer

The lexer stage does not understand the DSL itself, it only maps specific character sequences into identifiers.

This allows Icinga to detect the beginning of a string with ", reading the following characters and determining the end of the string with again ".

Other parts covered by the lexer a escape sequences insides a string, e.g. "\"abc".

The lexer also identifiers logical operators, e.g. & or in, specific keywords like object, import, etc. and comment blocks.

Please check lib/config/config_lexer.ll for details.

Icinga uses Flex in the first stage.

Flex (The Fast Lexical Analyzer)

Flex is a fast lexical analyser generator. It is a tool for generating programs that perform pattern-matching on text. Flex is a free (but non-GNU) implementation of the original Unix lex program.

Parser

The parser stage puts the identifiers from the lexer into more context with flow control and sequences.

The following comparison is parsed into a left term, an operator and a right term.

x > 5

The DSL contains many elements which require a specific order, and sometimes only a left term for example.

The parser also takes care of parsing an object declaration for example. It already knows from the lexer that object marks the beginning of an object. It then expects a type string afterwards, and the object name - which can be either a string with double quotes or a previously defined constant.

An opening bracket { in this specific context starts the object scope, which also is stored for later scope specific variable access.

If there's an apply rule defined, this follows the same principle. The config parser detects the scope of an apply rule and generates Icinga 2 C++ code for the parsed string tokens.

assign where host.vars.sla == "24x7"

is parsed into an assign token identifier, and the string expression is compiled into a new ApplyExpression object.

The flow control inside the parser ensures that for example ignore where can only be defined when a previous assign where was given - or when inside an apply for rule.

Another example are specific object types which allow assign expression, specifically group objects. Others objects must throw a configuration error.

Please check lib/config/config_parser.yy for more details, and the language reference chapter for documented DSL keywords and sequences.

Icinga uses Bison as parser generator which reads a specification of a context-free language, warns about any parsing ambiguities, and generates a parser in C++ which reads sequences of tokens and decides whether the sequence conforms to the syntax specified by the grammar.

Compiler

The config compiler initializes the scanner inside the lexer stage.

The configuration files are parsed into memory from inside the daemon CLI command which invokes the config validation in ValidateConfigFiles(). This compiles the files into an AST expression which is executed.

At this stage, the expressions generate so-called "config items" which are a pre-stage of the later compiled object.

ConfigItem::CommitItems takes care of committing the items, and doing a rollback on failure. It also checks against matching apply rules from the previous run and generates statistics about the objects which can be seen by the config validation.

ConfigItem::CommitNewItems collects the registered types and items, and checks for a specific required order, e.g. a service object needs a host object first.

The following stages happen then:

• Commit: A workqueue then commits the items in a parallel fashion for this specific type. The object gets its name, and the AST expression is executed. It is then registered into the item into m_Object as reference.
• OnAllConfigLoaded: Special signal for each object to pre-load required object attributes, resolve group membership, initialize functions and timers.
• CreateChildObjects: Run apply rules for this specific type.
• CommitNewItems: Apply rules may generate new config items, this is to ensure that they again run through the stages.

Note that the items are now committed and the configuration is validated and loaded into memory. The final config objects are not yet activated though.

This only happens after the validation, when the application is about to be run with ConfigItem::ActivateItems.

Each item has an object created in m_Object which is checked in a loop. Again, the dependency order of activated objects is important here, e.g. logger features come first, then config objects and last the checker, api, etc. features. This is done by sorting the objects based on their type specific activation priority.

The following signals are triggered in the stages:

• PreActivate: Setting the active flag for the config object.
• Activate: Calls Start() on the object, sets the local HA authority and notifies subscribers that this object is now activated (e.g. for config updates in the DB backend).

Features

Features are implemented in specific libraries and can be enabled using CLI commands.

Features either write specific data or receive data.

Examples for writing data: DB IDO, Graphite, InfluxDB. GELF, etc. Examples for receiving data: REST API, etc.

The implementation of features makes use of existing libraries and functionality. This makes the code more abstract, but shorter and easier to read.

Features register callback functions on specific events they want to handle. For example the GraphiteWriter feature subscribes to new CheckResult events.

Each time Icinga 2 receives and processes a new check result, this event is triggered and forwarded to all subscribers.

The GraphiteWriter feature calls the registered function and processes the received data. Features which connect Icinga 2 to external interfaces normally parse and reformat the received data into an applicable format.

Since this check result signal is blocking, many of the features include a work queue with asynchronous task handling.

The GraphiteWriter uses a TCP socket to communicate with the carbon cache daemon of Graphite. The InfluxDBWriter is instead writing bulk metric messages to InfluxDB's HTTP API, similar to Elasticsearch.

Check Scheduler

The check scheduler starts a thread which loops forever. It waits for check events being inserted into m_IdleCheckables.

If the current pending check event number is larger than the configured max concurrent checks, the thread waits up until it there's slots again.

In addition, further checks on enabled checks, check periods, etc. are performed. Once all conditions have passed, the next check timestamp is calculated and updated. This also is the timestamp where Icinga expects a new check result ("freshness check").

The object is removed from idle checkables, and inserted into the pending checkables list. This can be seen via REST API metrics for the checker component feature as well.

The actual check execution happens asynchronously using the application's thread pool.

Once the check returns, it is removed from pending checkables and again inserted into idle checkables. This ensures that the scheduler takes this checkable event into account in the next iteration.

Start

When checkable objects get activated during the startup phase, the checker feature registers a handler for this event. This is due to the fact that the checker feature is fully optional, and e.g. not used on command endpoint clients.

Whenever such an object activation signal is triggered, Icinga 2 checks whether it is authoritative for this object. This means that inside an HA enabled zone with two endpoints, only non-paused checkable objects are actively inserted into the idle checkable list for the check scheduler.

Initial Check

When a new checkable object (host or service) is initially added to the configuration, Icinga 2 performs the following during startup:

• Checkable::Start() is called and calculates the first check time
• With a spread delta, the next check time is actually set.

If the next check should happen within a time frame of 60 seconds, Icinga 2 calculates a delta from a random value. The minimum of check_interval and 60 seconds is used as basis, multiplied with a random value between 0 and 1.

In the best case, this check gets immediately executed after application start. The worst case scenario is that the check is scheduled 60 seconds after start the latest.

The reasons for delaying and spreading checks during startup is that the application typically needs more resources at this time (cluster connections, feature warmup, initial syncs, etc.). Immediate check execution with thousands of checks could lead into performance problems, and additional events for each received check results.

Therefore the initial check window is 60 seconds on application startup, random seed for all checkables. This is not predictable over multiple restarts for specific checkable objects, the delta changes every time.

Scheduling Offset

There's a high chance that many checkable objects get executed at the same time and interval after startup. The initial scheduling spreads that a little, but Icinga 2 also attempts to ensure to keep fixed intervals, even with high check latency.

During startup, Icinga 2 calculates the scheduling offset from a random number:

• Checkable::Checkable() calls SetSchedulingOffset() with Utility::Random()
• The offset is a pseudo-random integral value between 0 and RAND_MAX.

Whenever the next check time is updated with Checkable::UpdateNextCheck(), the scheduling offset is taken into account.

Depending on the state type (SOFT or HARD), either the retry_interval or check_interval is used. If the interval is greater than 1 second, the time adjustment is calculated in the following way:

now * 100 + offset divided by interval * 100, using the remainder (that's what fmod() is for) and dividing this again onto base 100.

Example: offset is 6500, interval 300, now is 1542190472.

1542190472 * 100 + 6500 = 154219053714
300 * 100 = 30000
154219053714 / 30000 = 5140635.1238

(5140635.1238 - 5140635.0) * 30000 = 3714
3714 / 100 = 37.14

37.15 seconds as an offset would be far too much, so this is again used as a calculation divider for the real offset with the base of 5 times the actual interval.

Again, the remainder is calculated from the offset and interval * 5. This is divided onto base 100 again, with an additional 0.5 seconds delay.

Example: offset is 6500, interval 300.

6500 / 300 = 21.666666666666667
(21.666666666666667 - 21.0) * 300 = 200
200 / 100 = 2
2 + 0.5 = 2.5

The minimum value between the first adjustment and the second offset calculation based on the interval is taken, in the above example 2.5 wins.

The actual next check time substracts the adjusted time from the future interval addition to provide a more widespread scheduling time among all checkable objects.

nextCheck = now - adj + interval

You may ask, what other values can happen with this offset calculation. Consider calculating more examples with different interval settings.

Example: offset is 34567, interval 60, now is 1542190472.

1542190472 * 100 + 34567 = 154219081767
60 * 100 = 6000
154219081767 / 6000 = 25703180.2945
(25703180.2945 - 25703180.0) * 6000 / 100 = 17.67

34567 / 60 = 576.116666666666667
(576.116666666666667 - 576.0) * 60 / 100 + 0.5 = 1.2

1m interval starts at now + 1.2s.

Example: offset is 12345, interval 86400, now is 1542190472.

1542190472 * 100 + 12345 = 154219059545
86400 * 100 = 8640000
154219059545 / 8640000 = 17849.428188078703704
(17849.428188078703704 - 17849) * 8640000 = 3699545
3699545 / 100 = 36995.45

12345 / 86400 = 0.142881944444444
0.142881944444444 * 86400 / 100 + 0.5 = 123.95

1d interval starts at now + 2m4s.

Note

In case you have a better algorithm at hand, feel free to discuss this in a PR on GitHub. It needs to fulfill two things: 1) spread and shuffle execution times on each next_check update 2) not too narrowed window for both long and short intervals Application startup and initial checks need to be handled with care in a slightly different fashion.

When SetNextCheck() is called, there are signals registered. One of them sits inside the CheckerComponent class whose handler CheckerComponent::NextCheckChangedHandler() deletes/inserts the next check event from the scheduling queue. This basically is a list with multiple indexes with the keys for scheduling info and the object.

Check Latency and Execution Time

Each check command execution logs the start and end time where Icinga 2 (and the end user) is able to calculate the plugin execution time from it.

GetExecutionEnd() - GetExecutionStart()

The higher the execution time, the higher the command timeout must be set. Furthermore users and developers are encouraged to look into plugin optimizations to minimize the execution time. Sometimes it is better to let an external daemon/script do the checks and feed them back via REST API.

Icinga 2 stores the scheduled start and end time for a check. If the actual check execution time differs from the scheduled time, e.g. due to performance problems or limited execution slots (concurrent checks), this value is stored and computed from inside the check result.

The difference between the two deltas is called check latency.

(GetScheduleEnd() - GetScheduleStart()) - CalculateExecutionTime()

Cluster

Communication

Icinga 2 uses its own certificate authority (CA) by default. The public and private CA keys can be generated on the signing master.

Each node certificate must be signed by the private CA key.

Note: The following description uses parent node and child node. This also applies to nodes in the same cluster zone.

During the connection attempt, an SSL handshake is performed. If the public certificate of a child node is not signed by the same CA, the child node is not trusted and the connection will be closed.

If the SSL handshake succeeds, the parent node reads the certificate's common name (CN) of the child node and looks for a local Endpoint object name configuration.

If there is no Endpoint object found, further communication (runtime and config sync, etc.) is terminated.

The child node also checks the CN from the parent node's public certificate. If the child node does not find any local Endpoint object name configuration, it will not trust the parent node.

Both checks prevent accepting cluster messages from an untrusted source endpoint.

If an Endpoint match was found, there is one additional security mechanism in place: Endpoints belong to a Zone hierarchy.

Several cluster messages can only be sent "top down", others like check results are allowed being sent from the child to the parent node.

Once this check succeeds the cluster messages are exchanged and processed.

CSR Signing

In order to make things easier, Icinga 2 provides built-in methods to allow child nodes to request a signed certificate from the signing master.

Icinga 2 v2.8 introduces the possibility to request certificates from indirectly connected nodes. This is required for multi level cluster environments with masters, satellites and clients.

CSR Signing in general starts with the master setup. This step ensures that the master is in a working CSR signing state with:

• public and private CA key in /var/lib/icinga2/ca
• private TicketSalt constant defined inside the api feature
• Cluster communication is ready and Icinga 2 listens on port 5665

The child node setup which is run with CLI commands will now attempt to connect to the parent node. This is not necessarily the signing master instance, but could also be a parent satellite node.

During this process the child node asks the user to verify the parent node's public certificate to prevent MITM attacks.

There are two methods to request signed certificates:

• Add the ticket into the request. This ticket was generated on the master beforehand and contains hashed details for which client it has been created. The signing master uses this information to automatically sign the certificate request.

• Do not add a ticket into the request. It will be sent to the signing master which stores the pending request. Manual user interaction with CLI commands is necessary to sign the request.

The certificate request is sent as pki::RequestCertificate cluster message to the parent node.

If the parent node is not the signing master, it stores the request in /var/lib/icinga2/certificate-requests and forwards the cluster message to its parent node.

Once the message arrives on the signing master, it first verifies that the sent certificate request is valid. This is to prevent unwanted errors or modified requests from the "proxy" node.

After verification, the signing master checks if the request contains a valid signing ticket. It hashes the certificate's common name and compares the value to the received ticket number.

If the ticket is valid, the certificate request is immediately signed with CA key. The request is sent back to the client inside a pki::UpdateCertificate cluster message.

If the child node was not the certificate request origin, it only updates the cached request for the child node and send another cluster message down to its child node (e.g. from a satellite to a client).

If no ticket was specified, the signing master waits until the ca sign CLI command manually signed the certificate.

Note

Push notifications for manual request signing is not yet implemented (TODO).

Once the child node reconnects it synchronizes all signed certificate requests. This takes some minutes and requires all nodes to reconnect to each other.

CSR Signing: Clients without parent connection

There is an additional scenario: The setup on a child node does not necessarily need a connection to the parent node.

This mode leaves the node in a semi-configured state. You need to manually copy the master's public CA key into /var/lib/icinga2/certs/ca.crt on the client before starting Icinga 2.

The parent node needs to actively connect to the child node. Once this connections succeeds, the child node will actively request a signed certificate.

The update procedure works the same way as above.

High Availability

General high availability is automatically enabled between two endpoints in the same cluster zone.

This requires the same configuration and enabled features on both nodes.

HA zone members trust each other and share event updates as cluster messages. This includes for example check results, next check timestamp updates, acknowledgements or notifications.

This ensures that both nodes are synchronized. If one node goes away, the remaining node takes over and continues as normal.

High Availability: Object Authority

Cluster nodes automatically determine the authority for configuration objects. By default, all config objects are set to HARunEverywhere and as such the object authority is true for any config object on any instance.

Specific objects can override and influence this setting, e.g. with HARunOnce instead prior to config object activation.

This is done when the daemon starts and in a regular interval inside the ApiListener class, specifically calling ApiListener::UpdateObjectAuthority().

The algorithm works like this:

• Determine whether this instance is assigned to a local zone and endpoint.
• Collects all endpoints in this zone if they are connected.
• If there's two endpoints, but only us seeing ourselves and the application start is less than 60 seconds in the past, do nothing (wait for cluster reconnect to take place, grace period).
• Sort the collected endpoints by name.
• Iterate over all config types and their respective objects
• Ignore !active objects
• Ignore objects which are !HARunOnce. This means, they can run multiple times in a zone and don't need an authority update.
• If this instance doesn't have a local zone, set authority to true. This is for non-clustered standalone environments where everything belongs to this instance.
• Calculate the object authority based on the connected endpoint names.
• Set the authority (true or false)

The object authority calculation works "offline" without any message exchange. Each instance alculates the SDBM hash of the config object name, puts that in contrast modulo the connected endpoints size. This index is used to lookup the corresponding endpoint in the connected endpoints array, including the local endpoint. Whether the local endpoint is equal to the selected endpoint, or not, this sets the authority to true or false.

authority = endpoints[Utility::SDBM(object->GetName()) % endpoints.size()] == my_endpoint;

ConfigObject::SetAuthority(bool authority) triggers the following events:

• Authority is true and object now paused: Resume the object and set paused to false.
• Authority is false, object not paused: Pause the object and set paused to true.

This results in activated but paused objects on one endpoint. You can verify that by querying the paused attribute for all objects via REST API or debug console on both endpoints.

Endpoints inside a HA zone calculate the object authority independent from each other. This object authority is important for selected features explained below.

Since features are configuration objects too, you must ensure that all nodes inside the HA zone share the same enabled features. If configured otherwise, one might have a checker feature on the left node, nothing on the right node. This leads to late check results because one half is not executed by the right node which holds half of the object authorities.

By default, features are enabled to "Run-Everywhere". Specific features which support HA awareness, provide the enable_ha configuration attribute. When enable_ha is set to true (usually the default), "Run-Once" is set and the feature pauses on one side.

vim /etc/icinga2/features-enabled/graphite.conf

object GraphiteWriter "graphite" {
  ...
  enable_ha = true
}

Once such a feature is paused, there won't be any more event handling, e.g. the Elasticsearch feature won't process any checkresults nor write to the Elasticsearch REST API.

When the cluster connection drops, the feature configuration object is updated with the new object authority by the ApiListener timer and resumes its operation. You can see that by grepping the log file for resumed and paused.

[2018-10-24 13:28:28 +0200] information/GraphiteWriter: 'g-ha' paused.

[2018-10-24 13:28:28 +0200] information/GraphiteWriter: 'g-ha' resumed.

Specific features with HA capabilities are explained below.

High Availability: Checker

The checker feature only executes checks for Checkable objects (Host, Service) where it is authoritative.

That way each node only executes checks for a segment of the overall configuration objects.

The cluster message routing ensures that all check results are synchronized to nodes which are not authoritative for this configuration object.

High Availability: Notifications

The notification feature only sends notifications for Notification objects where it is authoritative.

That way each node only executes notifications for a segment of all notification objects.

Notified users and other event details are synchronized throughout the cluster. This is required if for example the DB IDO feature is active on the other node.

High Availability: DB IDO

If you don't have HA enabled for the IDO feature, both nodes will write their status and historical data to their own separate database backends.

In order to avoid data separation and a split view (each node would require its own Icinga Web 2 installation on top), the high availability option was added to the DB IDO feature. This is enabled by default with the enable_ha setting.

This requires a central database backend. Best practice is to use a MySQL cluster with a virtual IP.

Both Icinga 2 nodes require the connection and credential details configured in their DB IDO feature.

During startup Icinga 2 calculates whether the feature configuration object is authoritative on this node or not. The order is an alpha-numeric comparison, e.g. if you have master1 and master2, Icinga 2 will enable the DB IDO feature on master2 by default.

If the connection between endpoints drops, the object authority is re-calculated.

In order to prevent data duplication in a split-brain scenario where both nodes would write into the same database, there is another safety mechanism in place.

The split-brain decision which node will write to the database is calculated from a quorum inside the programstatus table. Each node verifies whether the endpoint_name column is not itself on database connect. In addition to that the DB IDO feature compares the last_update_time column against the current timestamp plus the configured failover_timeout offset.

That way only one active DB IDO feature writes to the database, even if they are not currently connected in a cluster zone. This prevents data duplication in historical tables.

Health Checks

cluster-zone

This built-in check provides the possibility to check for connectivity between zones.

If you for example need to know whether the master zone is connected and processing messages with the child zone called satellite in this example, you can configure the cluster-zone check as new service on all master zone hosts.

vim /etc/zones.d/master/host1.conf

object Service "cluster-zone-satellite" {
  check_command = "cluster-zone"
  host_name = "host1"

  vars.cluster_zone = "satellite"
}

The check itself changes to NOT-OK if one or more child endpoints in the child zone are not connected to parent zone endpoints.

In addition to the overall connectivity check, the log lag is calculated based on the to-be-sent replay log. Each instance stores that for its configured endpoint objects.

This health check iterates over the target zone (cluster_zone) and their endpoints.

The log lag is greater than zero if

• the replay log synchronization is in progress and not yet finished or
• the endpoint is not connected, and no replay log sync happened (obviously).

The final log lag value is the worst value detected. If satellite1 has a log lag of 1.5 and satellite2 only has 0.5, the computed value will be 1.5..

You can control the check state by using optional warning and critical thresholds for the log lag value.

If this service exists multiple times, e.g. for each master host object, the log lag may differ based on the execution time. This happens for example on restart of an instance when the log replay is in progress and a health check is executed at different times. If the endpoint is not connected, both master instances may have saved a different log replay position from the last synchronisation.

The lag value is returned as performance metric key slave_lag.

Icinga 2 v2.9+ adds more performance metrics for these values:

• last_messages_sent and last_messages_received as UNIX timestamp
• sum_messages_sent_per_second and sum_messages_received_per_second
• sum_bytes_sent_per_second and sum_bytes_received_per_second

TLS Network IO

TLS Connection Handling

TLS-Handshake timeouts occur if the server is busy with reconnect handling and other tasks which run in isolated threads. Icinga 2 uses threads in many ways, e.g. for timers to wake them up, wait for check results, etc.

In terms of the cluster communication, the following flow applies.

Master Connects

• The master initializes the connection in a loop through all known zones it should connect to, extracting the endpoints and their host/port attribute.
• This calls AddConnection() whereas a Tcp::Connect() is called to create a TCP socket.
• A new thread is spawned for future connection handling, this binds ApiListener::NewClientHandler().
• On top of the TCP socket, a new TLS stream is created.
• The master performs a TLS->Handshake()
• Certificates are verified and the endpoint name is compared to the CN.

Clients Processes Connection

• The client listens for new incoming connections as 'TCP server' pattern inside ListenerThreadProc() with an endless loop.
• Once a new connection is detected, TCP->Accept() performs the initial socket establishment.
• A new thread is spawned for future connection handling, this binds ApiListener::NewClientHandler(), Role being Server.
• On top of the TCP socket, a new TLS stream is created.
• The client performs a TLS->Handshake().

Data Transmission between Server and Client Role

Once the TLS handshake and certificate verification is completed, the role is either Client or Server.

• Client: Send "Hello" message.
• Server: TLS->WaitForData() waits for incoming messages from the remote client.

Client in this case is the instance which initiated the connection. If the master is doing this, the Icinga 2 client/agent acts as "server" which accepts incoming connections.

Asynchronous Socket IO

Everything runs through TLS, we don't use any "raw" connections nor plain message handling.

The TLS handshake and further read/write operations are not performed in a synchronous fashion in the new client's thread. Instead, all clients share an asynchronous "event pool".

The TlsStream constructor registers a new SocketEvent by calling its constructor. It binds the previously created TCP socket and itself into the created SocketEvent object.

SocketEvent::InitializeEngine() takes care of whether to use epoll (Linux) or poll (BSD, Unix, Windows) as preferred socket poll engine. epoll has proven to be faster on Linux systems.

The selected engine is stored as l_SocketIOEngine and later Start() ensures to do the following:

• Use a fixed number for creating IO threads.
• Create a dumb_socketpair which basically is a pipe from in->out and multiplexes the TCP socket into a local Unix socket. This removes the complexity and slowlyness of the kernel dealing with the TCP stack and new events.
• InitializeThread() prepares epoll with epoll_create, socket descriptors and event mapping for later wakeup.
• Each event FD has its own "worker event thread" which deals with incoming data, called ThreadProc as endless loop.

By default, there are 8 of these worker threads.

In the ThreadProc loop, the following happens:

• epoll_wait gets called and provides an event whether new data is ready (via socket IO from the Kernel).
• The event created with epoll_event holds the .fd.data attribute which references the multiplexed event FD (and therefore tcp socket FD).
• All events in this cycle are stored with their descriptors in a list.
• Once the epoll loop is finished, the collected events are processed and the socketevent descriptor (which is the TlsStream object) calls OnEvent().

On Socket Event State Machine

OnEvent implements the "state machine" depending on the current desired action. By default, this is TlsActionNone.

Once TlsStream->Handshake() is called, this initializes the current action to TlsActionHandshake and performs SSL_do_handshake(). This function returns > 0 when successful, anything below needs to be dealt separately.

If the handshake was successful, the registered condition variable m_CV gets signalled and the thread waiting for the handshake in TlsStream->Handshake() wakes up and continues within the ApiListener::NewClientHandler() function.

Once the handshake is completed, current action is changed to either TlsActionRead or TlsActionWrite. This happens in the beginning of the state machine when there is no action selected yet.

• Read: Received events indicate POLLIN (or POLLERR/POLLHUP as error, but normally mean "read").
• Write: The send buffer of the TLS stream is greater 0 bytes, and the received events allow POLLOUT on the event socket.
• Nothing matched: Change the event sockets to POLLIN ("read"), and return, waiting for the next event.

This also depends on the returned error codes of the SSL interface functions. Whenever SSL_WANT_READ occurs, the event polling needs be changed to use POLLIN, vice versa for SSL_WANT_WRITE and POLLOUT.

In the scenario where the master actively connects to the clients, the client will wait for data and change the event sockets to Read once there's something coming on the sockets.

Action	Description
Read	Calls SSL_read() with a fixed buffer size of 64 KB. If rc > 0, the receive buffer of the TLS stream is filled and success indicated. This endless loop continues until a) SSL_pending() says no more data from remote b) Maximum bytes are read. If success is true, the condition variable notifies the thread in WaitForData to wake up.
Write	The send buffer of the TLS stream Peek()s the first 64KB and calls SSL_write() to send them over the socket. The returned value is the number of bytes written, this is adjusted within the send buffer in the Read() call (it also optimizes the memory usage).
Handshake	Calls SSL_do_handshake() and if successful, the condition variable wakes up the thread waiting for it in Handshake().

TLS Error Handling

TLS error code	Description
SSL_WANT_READ	The next event should read again, change events to POLLIN.
SSL_ERROR_WANT_WRITE	The next event should write, change events to POLLOUT.
SSL_ERROR_ZERO_RETURN	Nothing was returned, close the TLS stream and immediately return.
default	Extract the error code and log a fancy error for the user. Close the connection.

From this question:

With non-blocking sockets, SSL_WANT_READ means "wait for the socket to be readable, then call this function again."; conversely, SSL_WANT_WRITE means "wait for the socket to be writeable, then call this function again.". You can get either SSL_WANT_WRITE or SSL_WANT_READ from both an SSL_read() or SSL_write() call.

Successful TLS Actions

• Initialize the next TLS action to none. This re-evaluates the conditions upon next event call.
• If the stream still contains data, adjust the socket events.
  • If the send buffer contains data, change events to POLLIN|POLLOUT.
  • Otherwise POLLIN to wait for data.
• Process data when the receive buffer has them available and we are actively handling events.
• If the TLS stream is supposed to shutdown, close everything including the TLS connection.

Data Processing

Once a stream has data available, it calls SignalDataAvailable(). This holds a condition variable which wakes up another thread in a handled which was previously registered, e.g. for JsonRpcConnection, HttpServerConnection or HttpClientConnection objects.

All of them read data from the stream and process the messages. At this point the string is available as JSON already and later decoded (e.g. Icinga data structures, as Dictionary).

General Design Patterns

Taken from https://www.ibm.com/developerworks/aix/library/au-libev/index.html

One of the biggest problems facing many server deployments, particularly web server deployments, is the ability to handle a large number of connections. Whether you are building cloud-based services to handle network traffic, distributing your application over IBM Amazon EC instances, or providing a high-performance component for your web site, you need to be able to handle a large number of simultaneous connections.

A good example is the recent move to more dynamic web applications, especially those using AJAX techniques. If you are deploying a system that allows many thousands of clients to update information directly within a web page, such as a system providing live monitoring of an event or issue, then the speed at which you can effectively serve the information is vital. In a grid or cloud situation, you might have permanent open connections from thousands of clients simultaneously, and you need to be able to serve the requests and responses to each client.

Before looking at how libevent and libev are able to handle multiple network connections, let's take a brief look at some of the traditional solutions for handling this type of connectivity.

### Handling multiple clients

There are a number of different traditional methods that handle multiple connections, but usually they result in an issue handling large quantities of connections, either because they use too much memory, too much CPU, or they reach an operating system limit of some kind.

The main solutions used are:

* Round-robin: The early systems use a simple solution of round-robin selection, simply iterating over a list of open network connections and determining whether there is any data to read. This is both slow (especially as the number of connections increases) and inefficient (since other connections may be sending requests and expecting responses while you are servicing the current one). The other connections have to wait while you iterate through each one. If you have 100 connections and only one has data, you still have to work through the other 99 to get to the one that needs servicing.
* poll, epoll, and variations: This uses a modification of the round-robin approach, using a structure to hold an array of each of the connections to be monitored, with a callback mechanism so that when data is identified on a network socket, the handling function is called. The problem with poll is that the size of the structure can be quite large, and modifying the structure as you add new network connections to the list can increase the load and affect performance.
* select: The select() function call uses a static structure, which had previously been hard-coded to a relatively small number (1024 connections), which makes it impractical for very large deployments.
There are other implementations on individual platforms (such as /dev/poll on Solaris, or kqueue on FreeBSD/NetBSD) that may perform better on their chosen OS, but they are not portable and don't necessarily resolve the upper level problems of handling requests.

All of the above solutions use a simple loop to wait and handle requests, before dispatching the request to a separate function to handle the actual network interaction. The key is that the loop and network sockets need a lot of management code to ensure that you are listening, updating, and controlling the different connections and interfaces.

An alternative method of handling many different connections is to make use of the multi-threading support in most modern kernels to listen and handle connections, opening a new thread for each connection. This shifts the responsibility back to the operating system directly but implies a relatively large overhead in terms of RAM and CPU, as each thread will need it's own execution space. And if each thread (ergo network connection) is busy, then the context switching to each thread can be significant. Finally, many kernels are not designed to handle such a large number of active threads.

Alternative Implementations and Libraries

While analysing Icinga 2's socket IO event handling, the libraries and implementations below have been collected too. This thread also sheds more light in modern programming techniques.

Our main "problem" with Icinga 2 are modern compilers supporting the full C++11 feature set. Recent analysis have proven that gcc on CentOS 6 or SLES11 are too old to use modern programming techniques or anything which implemens C++14 at least.

Given the below projects, we are also not fans of wrapping C interfaces into C++ code in case you want to look into possible patches.

One key thing for external code is license compatibility with GPLv2. Modified BSD and Boost can be pulled into the third-party/ directory, best header only and compiled into the Icinga 2 binary.

C

• libevent: http://www.wangafu.net/~nickm/libevent-book/TOC.html
• libev: https://www.ibm.com/developerworks/aix/library/au-libev/index.html
• libuv: http://libuv.org

C++

• Asio (standalone header only or as Boost library): http://think-async.com (the Boost Software license is compatible with GPLv2)
• Poco project: https://github.com/pocoproject/poco
• cpp-netlib: https://github.com/cpp-netlib/cpp-netlib
• evpp: https://github.com/Qihoo360/evpp