svn commit: r351456 - head/sys/amd64/amd64

[Conrad Meyer]

r351456 only loosened restrictions on some of the less common thread
types; it was accidentally necessary, but not sufficient.  351494,
351495, and 351496 (at least) are also necessary, once the issue was
identified.

Best,
Conrad

On Mon, Aug 26, 2019 at 9:25 AM John Baldwin <jhb at freebsd.org> wrote:
>
> On 8/24/19 1:43 PM, Konstantin Belousov wrote:
> > On Sat, Aug 24, 2019 at 11:47:52AM -0700, Conrad Meyer wrote:
> >> On Sat, Aug 24, 2019 at 9:15 AM Konstantin Belousov <kostikbel at gmail.com> wrote:
> >>>
> >>> On Sat, Aug 24, 2019 at 08:49:42AM -0700, Conrad Meyer wrote:
> >>>> Hi Konstantin,
> >>>>
> >>>> What is the motivation for this change?  The commit message doesn't
> >>>> really describe why it was done.
> >>>
> >>> Really it does. There is no point to request allocations for e.g.
> >>> doublefault stack to be at the local domain, because this stack is only
> >>> used once.  Doublefault is definitely a machine halt situation, it does
> >>> not matter if it generates inter-socket traffic to handle.
> >>>
> >>> Same for boot stacks, and for mce.
> >>>
> >>> The change avoids unnecessary constraints.
> >>
> >> Sure, but what is the harm of the unnecessary constraints?  Does this
> >> change fix an actual bug, or is it just a stylistic preference to
> >> avoid domain-specific allocations for infrequently used objects?
> > I am not sure about this being a stylistic preference.  We usually
> > write code to express the required actions.  I removed constraints
> > which did not added anything neither to code correctness nor to the
> > performance.
>
> Judging by the thread on current though, this fixes boot panics on
> machines with NUMA but CPUs that don't have local memory, correct?
> I think that's the thing Conrad is asking.
>
> --
> John Baldwin