Category: rsyslog

Past two day’s rsyslog work log:
2008-01-26
– improved SIGSEGV handler a bit (now we get a core dump)
– partially fixed bug that caused rsyslogd to stall processing enqueued messages
after turning off DA mode and before any new message were arrived (if a new
message arrived, everything went back to normal, so it was a temporary halt)
2008-01-27
– fixed the situation where message processing could be stalled for some period
after DA mode turn off
– reduced number of unnecessary wakeups of DA worker thread when high water mark
is not yet reached
– fixed a cosmetic issue in timeoutValue calculation (just used for displaying)
– fixed a bug when shutting down DA queue
– some more fixing and cleanup on the queue shutdown sequence
– fixed queue termination in case bSaveOnShutdown is 0
– implemented the $MainMsgQueueSaveOnShutdown config directive
– implemented the $MainMsgQueueWorkerThreadMinimumMessages config directive
– implemented the $MainMsgQueueTimeoutWorkerThreadShutdown config directive

Yesterday’s rsyslog work log:
2008-01-25
– restructured queue shutdown so that the queue timeout is properly applied before
terminatiing the queue
– disk-assisted queue mode finally begins to look good ;)
– added debug-support environment variable RSYSLOG_DEBUGLOG
– fixed worker timeout which was accidently not set (and thus 0)

Long time no rsyslog work log post. But this time its a rather long one:

2008-01-14
– removed no longer needed “Initialized” logic and variable
– worker shutdown sequence enhanced to try different ways to shut
down and terminate workers if none helps (this protects against
badly written output plugins which hold the queue for too long)
– implemented $MainMsgQueueTimeoutActionCompletion config directive
– implemented $MainMsgQueueTimeoutEnqueue config directive
– implemented $MainMsgQueueTimeoutShutdown config directive
– some cleanup
– removed $MainMsgQueueImmediateShutdown config directive and handling,
this has been superseeded by the Timeout params. Backward compatibility
is no concern, no version with that directive was ever released.
– released 3.10.2
– fixed a bug with standard template definitions – thanks to
varmojfekoj for spotting it
– implemented config file handlers for
* $MainMsgQueueHighWaterMark
* $MainMsgQueueLowWaterMark
* $MainMsgQueueDiscardMark
* $MainMsgQueueDiscardSeverity
but did NOT yet implement the functionality behind these directives!
– implemented $MainMsgQueueDiscardMark and $MainMsgQueueDiscardSeverity
(but serverity needs to be specified numerically for the time being)
– also implemented $MainMsgQueueDiscardMark logic on dequeue, changed
defaults
– implemented $MainMsgQueueDiscardMark == 0 –> disable Discard logic
– implemented $MainMsgQueueSize == 0 –> no limit on queue size
– begin disk assisted queue memory queue modes (not fully implemented
yet)
2008-01-15
– disk assisted queue works quite well, except for startup from disk queue
– changed startup of disk assisted mode to allow for higher concurrency,
most importantly allow the input to continue enqueue msgs while the
disk queue is initialized. This may help somewhat with UDP and other
lossy sources
– improved shutdown processing – in-memory queue is now drained to disk
– first shot at queue restore on startup, but could not finish before
I need to leave ;)
2008-01-16
– fixed a compile-time bug in release mode spotted by Michael Biebl
– queue is now able to restore persisted state on startup (but still some
fine tuning to be done)
– re-released 3.10.2 due to bug reported by Michael Biebl
– cleaned up queue disk startup
– implemented dynamic startup and shutdown of worker threads based on
current activity
– some cleanup and fixes
– some more cleanup and flagged places where we need to implement
DA-input-only mode
2008-01-17
– changed obj_t destructor interface
– worked on threading
– fixed sync issue on shutdown process if need to persist pure memory
queue to disk
– adapted DA logic to the ability to shut down all workers for inactivity
2008-01-18
– created an in-depth description of DA assisted queue mode
– snapshot of new thread coding – DA mode still does not work, but need
to save
2008-01-19
– seperated mutex for queue size management from those for queue thread
management
– some further cleanup on the mutexes
2008-01-20
– begun object model redesign for clearer design of queue class, causes
creation of wti and wtp classes
– initial creation of wti class (worker implementation missing)
2008-01-21
– continued implementing wti class
– added some mutex-support to srUtils
– initial creation of wti class (still under development)
2008-01-22
– continued to work on queue
– implemented debug interface
2008-01-23
– added mutex debug instrumentation
2008-01-24
– improved debugging support (faster, less memory used, function invocation
count is maintened)
– redesigned queue to utilize helper classes for threading support. This is finally
in a running state for regular (non disk-assisted) queues, with a minor nit
at shutdown. So I can finally commit the work again to CVS…

I’ve not posted anything the past days – not because I don’t do anything but because I am terrible busy with new features. The new massively multithreading engine is giving me a somewhat harder time than I expected (and I have to admit a design flaw which I am recovering from). Everything else is currently being pushed away by my desire to finish store-and-forward enabled actions by the end of the month. I am no longer sure I’ll meet that deadline, but at least I’ll try.

Updates to the blog will hopefully resume soon, as will new releases. I’ll post work logs soon.

So you now know at least what’s going on… BTW: I am looking for some test environments where store-and-forward and/or high performance (running on as many cores as possible) is in high demand.

The past day’s rsyslog work log:
2008-01-11
– undid part of yesterdays stage work – q worker 0 does not have management
chores, will use another solution (not needed yet)
– begun to permit queue to terminate without being drained
– fixed a starvation condition in queueWorker (pthread_yield() was needed)
could not be seen with any previously released code, came up during
new development
– added $MainMsgQueueImmediateShutdown config directive
– some name cleanup
– added non-circular file stream mode
– added some debug instrumentation to obj_t type, so that invalidly passed
objects can be detected (else we use the jump table and do not know why
everything messes up)
– file stream objects are now persistent on immediate queue shutdown (queue itself
is not yet fully persisted)
– support for object property bags added
– queue can now persist disk queue information on immediate shutdown
– added function to de-serialize a property bag (untested as other code is yet
missing)
– support for de-serializing strm objects added
– partial ability to read a disk queue back in (not completed, but would like
to save source for the weekend)
2008-01-13
– support for reading back persisted queue information completed
– added $MainMsgQueuePersistUpdateCount config file directive
– renamed $MainMsgQueuePersistUpdateCount config file directive to
$MainMsgQueueCheckpointInterval
– changed queue shutdown procedure a bit – stage work for queue shutdown
timeout setting

rsyslog work log for Jan, 9th and 10th (sorry, forgot to post yesterday):

2008-01-09
– implemented new GetSize() handler for config files
– implemented $MainMsgQueueMaxFileSize configuration directive
– implemented queue object method to set the file name prefix
– implemented $MainMsgQueueFilePrefix configuration directive
– created a generic stream class (for file access)
– changed queue class to use stream class
– some cleanup on object model
2008-01-10
– changed queue file name generation (to be more generic)
– made queue file names better readable
– added buffered output to stream class
– data record support added to stream output writer
– added write functions for several types to stream class
– changed objSerialize methods to work directly on the stream class
– prepared for 3.10.1 release
– changed some config parameters and some cleanup
– released 3.10.1
– fixed a bug that caused a segfault on startup when no $WorkDir directive
was specified in rsyslog.conf
– fixed a bug that caused a segfault on queues with types other than “disk”
– removed the no longer needed thread TermSyncTool
– re-released 3.10.1
– implemented strm object serializer (untested as the code required for test
is not yet present – hen/egg problem…)
– some cleanup
– implemented management function for worker thread 0 in order to change
queue workers dynamically — stage work

Yesterday’s rsyslog work log:

2008-01-08
– fixed doc bug — thanks to varmojfekoj for pointing it out
– fixed some memory leaks in new code — thanks to varmojfekoj for the patch
– implemented queue disk reader to switch to multiple files
– first implementation of “disk” queue mode finished. It still needs some
work and the deserializer needs also to be expanded, but the queue at
least performs well now.
– fixed a race condition that could occur when input modules were terminated
– added –enable-mudflap ./configure option
– completed deserialization support in msg object (but not deserializer itself)
– completed object deserializer

rsyslog work log. I am currently totally focused on the queue (and thus not very responsive on mail and otherwise):

2008-01-07
– released 3.10.0
– performance-tuned stringbuf class
– implemented disk queue as far as I could without an object de-serializer
– implemented buffered read calls for the queue file
– implemented class type registry
– MsgSetProperty() implemented
– defined a property class
– implemented deserializer (needs some more work)

Here is the rsyslog work log for yesterday:

2008-01-06
– fixed a bug with integer conversion in srUtils.c
– changed some lib functions to work on long instead of int
to care for 64 bit platforms (just to be on the save side)
– worked a bit on object serialization
– cleaned up msg structure (interestingly, there were for example
two fields with identical meaning and iSyslogVersion was never
used ;))
– completed serializer for msg (but needs review)
– did a little bit performance cleanup
– worked on object header (now also contains the size)

If you followed my work logs or CVS updates, you’ve probably seen that I have worked quite a bit on rsyslog‘s threading. So I thought I share a view “design documents” that cover up the big picture.

Michael Biebel asked me for a few graphical representations of how the modules interact and what the message flow is. I am not a real good computer graphics guy, and an old-fashioned one. So I thought before I let you wait any longer, I share some of my hand sketches. They are not fancy, probably hard to read – but maybe still helpful. Find them below. A klick bring up the hires version, which is a bit less hard to read ;)

I’ll try to add better graphics and descriptions as soon as I find some time. But I have to admit that I currently have so many things on my mind that I’d like to code first. So it may take a short while.