[Sbcl-devel] triggering gc

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

* triggerring gc

As the code uses something like:

  if(!data->pending_handler) maybe_defer_handler(interrupt_maybe_gc_int,...)

triggerring gc is not reliable especially under heavy interrupt load (read:=
=20
sb-sprof). Furthermore, in a WITHOUT-INTERRUPTS gc is not triggered, becaus=
e=20
maybe_defer_handler defers it.

* delivery of SIG_STOP_FOR_GC

In a WITHOUT-INTERRUPTS block this signal is deferred, thus the gc cannot=20
start until an exit from it.

* WITHOUT-GCING

Thread A can already be stopping the world when thread B enters the=20
WITHOUT-GCING body and since the SIG_STOP_FOR_GC is not blocked thread B is=
=20
stopped and gc starts.


Assuming these are undesirable (including WITHOUT-INTERRUPTS deferring gc (=
?))=20
my current plan is to add gc_pending, sig_stop_for_gc_pending fields to=20
interrupt_data and:

=2D make WITHOUT-INTERRUPTS let gcs and SIG_STOP_FOR_GCs through

=2D make WITHOUT-GCING and pseudo atomic block/defer SIG_STOP_FOR_GC

=2D give priority to gc over signal handling (?!), if there is an interrupt=
 and=20
a gc/SIG_STOP_FOR_GC pending do the gc first

Currently, I'm reading up on the fabulous and rich history of gencgc.c, but=
 in=20
the meantime let me know if you have a simple(r) scheme for this.

Cheers, G=E1bor