Interrupt in PARI's pari_malloc() hangs Sage

[vbraun]

This fails quite frequently:

sage: 0.binomial(1) ## line 6262 ##
0
sage: 0.binomial(-1) ## line 6264 ##
0
sage: 13.binomial(2r) ## line 6266 ##
78
sage: alarm(0.5); (2^100).binomial(2^22, algorithm='mpir') ## line 6271 ##
sage: alarm(0.5); (2^100).binomial(2^22, algorithm='pari') ## line 6275 ##

**********************************************************************
----------------------------------------------------------------------
sage -t --long src/sage/rings/integer.pyx  # Timed out (and interrupt failed)
----------------------------------------------------------------------

This is a consequence of a recent PARI upgrade, because malloc() is used in a place where it wasn't used before. The cause is exactly the same as #16850.

For PARI, we should be able to fix this using PARI_SIGINT_block which is like sig_block() but for the PARI library.

[jdemeyer]

It's the PARI library interface, so #10476 cannot be the cause. My bet would something in PARI itself.

[jdemeyer]

New commits:

​f16032b

Support PARI_SIGINT_block in interrupt framework

[vdelecroix]

I am infinitely happy that you found out a solution for this annoying interrupt! Though, I do not like the fact that now the interrupt code depends explicitely on pari. In an ideal world, all this interrupt buisness should go in upstream Cython, don't you think? Do you have an idea whether it would be possible to have something modular for the interrupt code?

[jdemeyer]

Replying to vdelecroix:

Though, I do not like the fact that now the interrupt code depends explicitely on pari.

I understand your sentiment, but at least it fixes a bug which I don't know how to fix otherwise. You should also understand that I never claimed that cysignals is 100% bullet-proof. We try to fix most use cases but now and then something breaks (see #24986 for example).

In an ideal world, all this interrupt buisness should go in upstream Cython, don't you think?

Moving this to upstream Cython won't magically fix the PARI problem.

Besides, I don't think that there is much reason to push it to Cython now that we have cysignals.