Writing Rsyslog Output Plugins

This page is the begin of some developer documentation for writing output plugins. Doing so is quite easy (and that was a design goal), but there currently is only sparse documentation on the process available. I was tempted NOT to write this guide here because I know I will most probably not be able to write a complete guide.

However, I finally concluded that it may be better to have same information and pointers than to have nothing.

Getting Started and Samples

The best to get started with rsyslog plugin development is by looking at existing plugins. All that start with "om" are output modules. That means they are primarily thought of being message sinks. In theory, however, output plugins may aggergate other functionality, too. Nobody has taken this route so far so if you would like to do that, it is highly suggested to post your plan on the rsyslog mailing list, first (so that we can offer advise).

The rsyslog distribution tarball contains two plugins that are extremely well targeted for getting started:

• omtemplate
• omstdout

In any case, you should also read the comments in ./runtime/module-template.h. Output plugins are build based on a large set of code-generating macros. These macros handle most of the plumbing needed by the interface. As long as no special callback to rsyslog is needed (it typically is not), an output plugin does not really need to be aware that it is executed by rsyslog. As a plug-in programmer, you can (in most cases) "code as usual". However, all macros and entry points need to be provided and thus reading the code comments in the files mentioned is highly suggested.

In short, the best idea is to start with a template. Let's assume you start by copying omtemplate. Then, the basic steps you need to do are:

• cp ./plugins/omtemplate ./plugins/your-plugin
• mv cd ./plugins/your-plugin
• vi Makefile.am, adjust to your-plugin
• mv omtemplate.c your-plugin.c
• cd ../..
• vi Makefile.am configure.ac
  search for omtemplate, copy and modify (follow comments)

Basically, this is all you need to do ... Well, except, of course, coding your plugin ;). For testing, you need rsyslog's debugging support. Some useful information is given in "troubleshooting rsyslog from the doc set.

Special Topics

Threading

Rsyslog uses massive parallel processing and multithreading. However, a plugin's entry points are guaranteed to be never called concurrently for the same action. That means your plugin must be able to be called concurrently by two or more threads, but you can be sure that for the same instance no concurrent calls happen. This is guaranteed by the interface specification and the rsyslog core guards against multiple concurrent calls. An instance, in simple words, is one that shares a single instanceData structure.

So as long as you do not mess around with global data, you do not need to think about multithreading (and can apply a purely sequential programming methodology).

Please note that duringt the configuraton parsing stage of execution, access to global variables for the configuration system is safe. In that stage, the core will only call sequentially into the plugin.

Getting Message Data

The doAction() entry point of your plugin is provided with messages to be processed. It will only be activated after filtering and all other conditions, so you do not need to apply any other conditional but can simply process the message.

Note that you do NOT receive the full internal representation of the message object. There are various (including historical) reasons for this and, among others, this is a design decision based on security.

Your plugin will only receive what the end user has configured in a $template statement. However, starting with 4.1.6, there are two ways of receiving the template content. The default mode, and in most cases sufficient and optimal, is to receive a single string with the expanded template. As I said, this is usually optimal, think about writing things to files, emailing content or forwarding it.

The important philosophy is that a plugin should never reformat any of such strings - that would either remove the user's ability to fully control message formats or it would lead to duplicating code that is already present in the core. If you need some formatting that is not yet present in the core, suggest it to the rsyslog project, best done by sending a patch ;), and we will try hard to get it into the core (so far, we could accept all such suggestions - no promise, though).

If a single string seems not suitable for your application, the plugin can also request access to the template components. The typical use case seems to be databases, where you would like to access properties via specific fields. With that mode, you receive a char ** array, where each array element points to one field from the template (from left to right). Fields start at arrray index 0 and a NULL pointer means you have reached the end of the array (the typical Unix "poor man's linked list in an array" design). Note, however, that each of the individual components is a string. It is not a date stamp, number or whatever, but a string. This is because rsyslog processes strings (from a high-level design look at it) and so this is the natural data type. Feel free to convert to whatever you need, but keep in mind that malformed packets may have lead to field contents you'd never expected...

If you like to use the array-based parameter passing method, think that it is only available in rsyslog 4.1.6 and above. If you can accept that your plugin will not be working with previous versions, you do not need to handle pre 4.1.6 cases. However, it would be "nice" if you shut down yourself in these cases - otherwise the older rsyslog core engine will pass you a string where you expect the array of pointers, what most probably results in a segfault. To check whether or not the core supports the functionality, you can use this code sequence:

BEGINmodInit()
	rsRetVal localRet;
	rsRetVal (*pomsrGetSupportedTplOpts)(unsigned long *pOpts);
	unsigned long opts;
	int bArrayPassingSupported;		/* does core support template passing as an array? */
CODESTARTmodInit
	*ipIFVersProvided = CURR_MOD_IF_VERSION; /* we only support the current interface specification */
CODEmodInit_QueryRegCFSLineHdlr
	/* check if the rsyslog core supports parameter passing code */
	bArrayPassingSupported = 0;
	localRet = pHostQueryEtryPt((uchar*)"OMSRgetSupportedTplOpts", &pomsrGetSupportedTplOpts);
	if(localRet == RS_RET_OK) {
		/* found entry point, so let's see if core supports array passing */
		CHKiRet((*pomsrGetSupportedTplOpts)(&opts));
		if(opts & OMSR_TPL_AS_ARRAY)
			bArrayPassingSupported = 1;
	} else if(localRet != RS_RET_ENTRY_POINT_NOT_FOUND) {
		ABORT_FINALIZE(localRet); /* Something else went wrong, what is not acceptable */
	}
	DBGPRINTF("omstdout: array-passing is %ssupported by rsyslog core.\n", bArrayPassingSupported ? "" : "not ");

	if(!bArrayPassingSupported) {
		DBGPRINTF("rsyslog core too old, shutting down this plug-in\n");
		ABORT_FINALIZE(RS_RET_ERR);
	}

The code first checks if the core supports the OMSRgetSupportedTplOpts() API (which is also not present in all versions!) and, if so, queries the core if the OMSR_TPL_AS_ARRAY mode is supported. If either does not exits, the core is too old for this functionality. The sample snippet above then shuts down, but a plugin may instead just do things different. In omstdout, you can see how a plugin may deal with the situation.

In any case, it is recommended that at least a graceful shutdown is made and the array-passing capability not blindly be used. In such cases, we can not guard the plugin from segfaulting and if the plugin (as currently always) is run within rsyslog's process space, that results in a segfault for rsyslog. So do not do this.

Batching of Messages

Starting with rsyslog 4.3.x, batching of output messages is supported. Previously, only a single-message interface was supported.

With the single message plugin interface, each message is passed via a separate call to the plugin. Most importantly, the rsyslog engine assumes that each call to the plugin is a complete transaction and as such assumes that messages be properly commited after the plugin returns to the engine.

With the batching interface, rsyslog employs something along the line of "transactions". Obviously, the rsyslog core can not make non-transactional outputs to be fully transactional. But what it can is support that the output tells the core which messages have been commited by the output and which not yet. The core can than take care of those uncommited messages when problems occur. For example, if a plugin has received 50 messages but not yet told the core that it commited them, and then returns an error state, the core assumes that all these 50 messages were not written to the output. The core then requeues all 50 messages and does the usual retry processing. Once the output plugin tells the core that it is ready again to accept messages, the rsyslog core will provide it with these 50 not yet commited messages again (actually, at this point, the rsyslog core no longer knows that it is re-submiting the messages). If, in contrary, the plugin had told rsyslog that 40 of these 50 messages were commited (before it failed), then only 10 would have been requeued and resubmitted.

In order to provide an efficient implementation, there are some (mild) constraints in that transactional model: first of all, rsyslog itself specifies the ultimate transaction boundaries. That is, it tells the plugin when a transaction begins and when it must finish. The plugin is free to commit messages in between, but it must commit all work done when the core tells it that the transaction ends. All messages passed in between a begin and end transaction notification are called a batch of messages. They are passed in one by one, just as without transaction support. Note that batch sizes are variable within the range of 1 to a user configured maximum limit. Most importantly, that means that plugins may receive batches of single messages, so they are required to commit each message individually. If the plugin tries to be "smarter" than the rsyslog engine and does not commit messages in those cases (for example), the plugin puts message stream integrity at risk: once rsyslog has notified the plugin of transacton end, it discards all messages as it considers them committed and save. If now something goes wrong, the rsyslog core does not try to recover lost messages (and keep in mind that "goes wrong" includes such uncontrollable things like connection loss to a database server). So it is highly recommended to fully abide to the plugin interface details, even though you may think you can do it better. The second reason for that is that the core engine will have configuration settings that enable the user to tune commit rate to their use-case specific needs. And, as a relief: why would rsyslog ever decide to use batches of one? There is a trivial case and that is when we have very low activity so that no queue of messages builds up, in which case it makes sense to commit work as it arrives. (As a side-note, there are some valid cases where a timeout-based commit feature makes sense. This is also under evaluation and, once decided, the core will offer an interface plus a way to preserve message stream integrity for properly-crafted plugins).

The second restriction is that if a plugin makes commits in between (what is perfectly legal) those commits must be in-order. So if a commit is made for message ten out of 50, this means that messages one to nine are also commited. It would be possible to remove this restriction, but we have decided to deliberately introduce it to simpify things.

Output Plugin Transaction Interface

In order to keep compatible with existing output plugins (and because it introduces no complexity), the transactional plugin interface is build on the traditional non-transactional one. Well... actually the traditional interface was transactional since its introduction, in the sense that each message was processed in its own transaction.

So the current doAction() entry point can be considered to have this structure (from the transactional interface point of view):

doAction()
    {
    beginTransaction()
    ProcessMessage()
    endTransaction()
    }
 

beginTransaction()
    {
    /* prepare for transaction */
    }

doAction()
    {
    ProcessMessage()
    /* maybe do partial commits */
    }

endTransaction()
    {
    /* commit (rest of) batch */
    }

INITChkCoreFeature(bCoreSupportsBatching, CORE_FEATURE_BATCHING);