Re: [RFC] Moving to Guile 2.2 and away from GUB

[David Kastrup]

David Kastrup <dak@gnu.org> writes:

> David Kastrup <dak@gnu.org> writes:
>
>> David Kastrup <dak@gnu.org> writes:
>>
>>> Jonas Hahnfeld via Discussions on LilyPond development
>>> <lilypond-devel@gnu.org> writes:
>>>
>>>> Am Donnerstag, dem 25.11.2021 um 08:40 +0100 schrieb Jonas Hahnfeld via
>>>> Discussions on LilyPond development:
>>>>> Am Mittwoch, dem 24.11.2021 um 23:01 +0100 schrieb Han-Wen Nienhuys:
>>>>> > On Wed, Nov 24, 2021 at 8:13 PM Jonas Hahnfeld via Discussions on
>>>>> > LilyPond development <lilypond-devel@gnu.org> wrote:
>>>>> > > Assuming no major problems are found, we could then move completely to
>>>>> > > Guile 2.2 and do releases without GUB. Both steps still require a bit
>>>>> > > of work; I'm relatively sure there is still a GC related issue causing
>>>>> > > (very rare) crashes with Guile 2.2
>>>>> > 
>>>>> > can you say more about this?
>>>>> 
>>>>> Not really: I remember weird crashes with dumps that were not
>>>>> immediately obvious.
>>>>
>>>> It's relatively reproducible, I get it every 2nd-3rd time for full
>>>> 'make doc's. Looking at the core dump, it's crashing in
>>>> System::derived_mark when marking all_elements_ because the std::vector
>>>> has garbage in its size field (it was probably freed). What I don't
>>>> understand is how that can happen since the only way of setting
>>>> all_elements_ is in System::init_elements where the originating SCM is
>>>> also protected...
>>>
>>> My guess is that std::vector is not freed but rather not initialised.
>>> That kind of thing is what the Preinit class is supposed to prevent, but
>>> System does not use it.
>>>
>>> Note also that all_elements_ is initialised to nullptr only _after_
>>> garbage collection has already been activated: the in-class
>>> initialisation syntax makes it easier to gloss over initialisation order
>>> (which is not changed).
>>>
>>> I'll try GC-proofing System.  Only time will tell whether this makes a
>>> difference and/or whether other problems remain.
>>
>> Somewhat embarrassingly, it would not appear that anything protects
>> *all_elements_ as such (only its elements are protected in
>> System::derived_mark).  That would be kind of a WTF even with Guile-1.8
>> .
>
> Well...
>
> System::init_elements has
>
>   set_object (this, "all-elements", scm_arr);
>
> So *all_elements_ is protected as part of being a Spanner.  As long as
> the spanner has not committed suicide (when it drops its object list).
> If it has committed suicide but not been collected yet, *all_elements_
> may be collected before the System itself.
>
> This is garbage.

Check out <https://gitlab.com/lilypond/lilypond/-/merge_requests/1025>
if you can trigger this with some frequency.  Also it may be worth
figuring out whether the crash occurs in connection with frenched
systems (which are explicitly removed before the end of typesetting).

-- 
David Kastrup