Patch: support callback on allocation overflow instead of calling abort()

[Colin Caine]

Hi Niels,

This sounds hopeful!


> To make any progress with adding a callback like this, we'd need to
> either:
>
> 1. Find out and document how to longjmp out from the callback safely.
>
> 2. Agree and document that when GMP invokes this callback, GMP state
> should be considered invalid. The process must not make any further
> calls to any GMP functions.
>
> And in the latter case, would that be satisfactory for Julia's use?

Yes, this latter case is sufficient for us.

Our users could catch the error and continue anyway, but we can
document, as you can, that this is risky.

On Tue, 6 Oct 2020, 15:59 Niels Möller, <nisse at lysator.liu.se> wrote:

> Colin Caine <cmcaine at gmail.com> writes:
>
> > Sorry, this patch was posted to this list last month, but I didn't see it
> > in the archives. See here for related discussion:
> > https://gmplib.org/list-archives/gmp-bugs/2020-September/004865.html
>
> I asked some questions (some off list), and haven't seen satisfactory
> answers. The thing is, it seems fairly safe to use the propsed callback
> to write a friendly error message, backtrace of involved julia
> functions, etc, and then exit the process.
>
> But my impression is that julia may also longjmp out and continue
> execution, and to me, that seems very brittle. You can have memory
> leaks, you may leave GMP data structures in an inconsistent state.
>
> To make any progress with adding a callback like this, we'd need to
> either:
>
> 1. Find out and document how to longjmp out from the callback safely.
>
> 2. Agree and document that when GMP invokes this callback, GMP state
>    should be considered invalid. The process must not make any further
>    calls to any GMP functions.
>
> And in the latter case, would that be satisfactory for Julia's use?
>
> There are other approaches to avoid these crashes, e.g., I think the
> recent emacs integration uses a (configurable) limit on bignum size, and
> will raise an emacs exception long before hitting GMP's limits, and the
> emacs process can go on running with no issues.
>
> Regards,
> /Niels
>
> --
> Niels Möller. PGP-encrypted email is preferred. Keyid 368C6677.
> Internet email is subject to wholesale government surveillance.
>