Re: Deadlock on fragmented memory?

[Taylor R Campbell <campbell+netbsd-tech-kern%mumble.net@localhost>]

> Date: Thu, 19 Oct 2017 13:16:09 -0700
> From: Chuck Silvers <chuq%chuq.com@localhost>
> 
> this latest patch sounds like a definite improvement to me.

kern_exec rev. 1.448

> as I recall, the original pre-pool exec_map code would free the
> physical args pages as part of freeing the kernel virtual space...
> if we're still using a pool here, will the pool cache all this?
> it would be nice to avoid ever paging exec args pages out to swap space.
> maybe we could mark the execargs pages as clean before freeing the
> execargs mapping back to the pool?  this is a minor issue compared to
> the current problem, I'd say proceed with your current patch for now
> and we can add any further refinements later.

With the correctness now addressed (I hope!), I will leave the task of
adding calls to uvm_map_clean(PGO_FREE) just before pool_put, and
testing what effects they have on performance in contrast to the
uvm_km_free calls that ad@ was trying to avoid in the first place, as
an exercise for a reader more dedicated to performance than I.

I, meanwhile, will be quietly disappointed in myself for having taken
six years to analyze and fix this bug, which is more than half the
time it has been around and nearly the entire time it has been in a
release!  (In my defence, I was pursuing other endeavours during those
six years as well.)