Hi all,
I'm working with Genode 10.05 on top of the OKL4 Microkernel. Recently, my application raised a division-by-zero. Shame on me - but I was very astonished when I found out that the root task (core) handles this error rather by chance than explicitly, for there's no real handler registered.
The error ipc the kernel sends to core, arrives in the sleep_forever function within core's main thread which sleeps in an ipc call expected never to return. On the division-by-zero, this ipc returns indeed, causing the ipc_client to indicate that its receive buffer be too small for the incoming message (containing the kernel-saved thread context). Upon that, core's main thread acknowledges the ipc to the kernel and performs the sleep_forever action again. The offending application thread is aborted by the kernel.
One might call that a proper handling, in fact, it's correct, but it would have been very nice to have a real handler inside of core giving a simple hint on the error. The current implementation makes it very hard to debug such errors (In fact, I had to debug into core to see where the 'copy_utcb_to_msgbuf' failed, then into the kernel itself to see what was actually signalled to core and then finally into my application's threads upon the very faint idea what to look for).
Furthermore, if the message buffer inside of core would have been big enough to take the kernel's ipc, even the initial 'hint' where to start debugging would have been missing ;)
Could you please make core at least a little more verbose at this point?
By the way I found out that the handler for an invalid-opcode is missing as well, but I don't know if there's a chance to catch those errors inside the root task, for they cause the kernel itself to enter the kernel debugger and hence halting the system (so rather an OKL4 issue).
Regards
Sven -- Sven Fülster
Hi Sven,
One might call that a proper handling, in fact, it's correct, but it would have been very nice to have a real handler inside of core giving a simple hint on the error. The current implementation makes it very hard to debug such errors (In fact, I had to debug into core to see where the 'copy_utcb_to_msgbuf' failed, then into the kernel itself to see what was actually signalled to core and then finally into my application's threads upon the very faint idea what to look for).
Furthermore, if the message buffer inside of core would have been big enough to take the kernel's ipc, even the initial 'hint' where to start debugging would have been missing ;)
Could you please make core at least a little more verbose at this point?
thanks a lot for this useful suggestion. In fact, currently we set L4_roottask as exception handler for each thread, which results in the behaviour you observed. We will investigate how to best remove this inconvenience. One short-term idea would be registering the pager thread as both exception handler and page-fault handler and let this thread print a message on each occurrence of an exception.
Cheers Norman
Hello Sven,
Could you please make core at least a little more verbose at this point?
coming back to the exception topic, here is a simple solution, which reports occurred exceptions as a log message. The patch hasn't made it into the last release but if it works for you, I'll commit it to the SVN.
Norman
-
mx＠...19... -
Norman Feske