Re: [Sbcl-devel] x86/unithread flaming death with araneida

[<me...@ho...>]

On Thursday 20 October 2005 11:41, Nikodemus Siivola wrote:
> This is just a quick note in case someone has a clue -- I'll gather
> up more details and make a test-case if I can when I find some extra
> time:
>
> Components: unithreaded SBCL 0.9.5.73 and .77, serve-event Aranaida
> direct
> from Alan's darcs, Linux 2.6.12 (in UML). No idea which parts of this
> monstrosity are necessary for reproduction yet, but an earlier
> Araneida did run happily on 0.9.5.something-early.
>
> Here's what little backtrace I have handy:
>
> Caught error: memory fault
> 0: (BACKTRACE 15 #<SYNONYM-STREAM :SYMBOL SB-SYS:*STDOUT* {9112271}>)
> 1: (ARANEIDA::HANDLER-DEBUGGER-HOOK
>      #<SB-KERNEL::MEMORY-FAULT-ERROR {AFA1CA1}>
>      #<unavailable argument>)
> 2: (INVOKE-DEBUGGER #<SB-KERNEL::MEMORY-FAULT-ERROR {AFA1CA1}>)
> 3: (ERROR SB-KERNEL::MEMORY-FAULT-ERROR)
> 4: (SB-KERNEL::MEMORY-FAULT-ERROR #<error printing object>)
> 5: ("foreign function: call_into_lisp" #<error printing object>)
> 6: ("foreign function: post_signal_tramp" #<error printing object>)
>
> Half the time it burns all the way to LDB, half the time it stays in
> lisp-land, half of _which_ it is effectively dead: FIND-SYMBOL*
> puking with TYPE-ERRORS from invalid indices when reading -- the
> indices being large negative numbers.

Damn. I've been debugging this on threaded builds for a week. See Alan=20
Shield's email. The components I used: threaded sbcl (.58+), Alan's=20
test araneida:

darcs get --tag=3Dslamtest-without-mutex=20
http://code.microarray.omrf.org/darcs/alan-araneida-testing/

It turned out to be a great test: it triggered the SA_NODEFER bug and a=20
threading/gc bug too. I've been meaning to reply to Alan but delayed it=20
thinking that I would be able to catch the remaining bugs which are=20
similar to what you describe.

I've found a nest of threading bugs in globaldb but they are hard to fix=20
without slowing down the compiler (and understanding the code for that=20
matter).

I suspect it's my interrupt changes or the gencgc stuff. I'll conduct a=20
binary search on the 0.9.5 series favouring .19 and .50.

G=E1bor

>
> Cheers,
>
>   -- Nikodemus              Schemer: "Buddha is small, clean, and
> serious."
>                    Lispnik: "Buddha is big, has hairy armpits, and
> laughs."