Lumiera
The new emerging NLE for GNU/Linux
State Final
Date Sat Jul 24 21:59:02 2010
Proposed by Christian Thaeter <ct@pipapo.org>

Handling of Signals in a multithreaded Application is little special, I define here how we going to implement this.

Description

By default in POSIX signals are send to whatever thread is running and handled there. This is quite unfortunate because a thread might be in some time constrained situation, hold some locks or have some special priority. The common way to handle this is blocking (most) signals in all threads except having one dedicated signal handling thread. Moreover it makes sense that the initial thread does this signal handling.

For Lumiera I propose to follow this practice and extend it a little by dedicate a thread to some management tasks. These are: * signal handling, see below. * resource management (resource-collector), waiting on a condition variable or message queue to execute actions. * watchdog for threads, not being part of the application schedulers but waking up periodically (infrequently, every so many seconds) and check if any thread got stuck (threads.h defines a deadline api which threads may use). We may add some flag to threads defining what to do with a given thread when it got stuck (emergency shutdown or just cancel the thread). Generally threads should not get stuck but we have to be prepared against rogue plugins and programming errors.

Signals which need to be handled

This are mostly proposals about how the application shall react on signals and comments about possible signals.

SIGTERM

Send on computer shutdown to all running apps. When running with GUI but we likely lost the Xserver connection before, this needs to be handled from the GUI. Nevertheless in any case (most importantly when running headless) we should do a fast application shutdown, no data/work should go lost, a checkpoint in the log is created. Some caveat might be that Lumiera has to sync a lot of data to disk. This means that usual timeouts from SIGTERM to SIGKILL as in nomal shutdown might be not sufficient, there is nothing we can do there. The user has to configure his system to extend this timeouts (alternative: see SIGUSR below).

SIGINT

This is the CTRL-C case from terminal, in most cases this means that a user wants to break the application immediately. We trigger an emergency shutdown. Recent actions are be logged already, so no work gets lost, but no checkpoint in the log gets created so one has to explicitly recover the interrupted state.

SIGBUS

Will be raised by I/O errors in mapped memory. This is a kindof exceptional signal which might be handled in induvidual threads. When the cause of the error is traceable then the job/thread worked on this data goes into a errorneous mode, else we can only do a emergency shutdown.

SIGFPE

Floating point exception, divison by zero or something similar. Might be allowed to be handled by each thread. In the global handler we may just ignore it or do an emergency shutdown. tbd.

SIGHUP

For daemons this signal is usually used to re-read configuration data. We shall do so too when running headless. When running with GUI this might be either act like SIGTERM or SIGINT. possibly this can be configureable.

SIGSEGV

Should not be handled, at the time a SEGV appears we are in a undefined state and anything we do may make things worse.

SIGUSR1

First user defined signal. Sync all data to disk, generate a checkpoint. The application may block until this is completed. This can be used in preparation of a shutdown or periodically to create some safe-points.

SIGUSR2

Second user defined signal. Produce diagnostics, to terminal and file.

SIGXCPU

CPU time limit exceeded. Emergency Shutdown.

SIGXFSZ

File size limit exceeded. Emergency Shutdown.

Tasks

We have appstate::maybeWait() which already does such a loop. It needs to be extended by the proposed things above.

Discussion

Pros

Cons

Alternatives

Rationale

This is rather common practice. I describe this here for Documentation purposes and to point out which details are not yet covered.

Comments

State → Final

ichthyo wants this to be a dedicated thread (own subsystem) instead running in the initial thread. Fixed this in the proposal above, this makes this accepted. Do 14 Apr 2011 03:40:41 CEST Christian Thaeter <ct@pipapo.org>