runtime: prevent calls to GOMAXPROCS while clearing P trace state

Currently, between the forEachP that ensures every P has a status in the
trace and readying dead Ps for the next generation, there's a big window
where GOMAXPROCS could run and change the number of Ps. In such
circumstances, P state will not get properly prepared for the next
generation, and we may fail to emit a status for a P.

This change addresses the problem by holding worldsema across both
operations. It also moves the status emission and state clearing to
before the generation transition because that makes the code *much*
clearer.

Currently we do both these things after the generation transition
targeting the next-next generation. The reason for this is to avoid an
extra forEachP (because we already handle P status in the STW for
enabling tracing to begin with) but approach is just plain harder to
reason about. Preparing the next generation before transitioning to it,
like we do for goroutines, is much clearer. Furthermore, we also need to
set up the dead P state even if we're stopping the trace, which would
mean duplicating code into both branches (if stopping the trace, then we
go non-preemptible, otherwise we do it under worldsema) which is also
less clear.

Note that with this change we no longer need to emit the P statuses
during the STW, which was probably the worse choice anyway, since
holding up the STW for an O(P) operation is worse than a forEachP we're
going to do anyway. So, this change does away with that too.

Fixes #76572.

Change-Id: Ie7908adff43a8a372cae432bcc2f4e0d6a87d7bc
Reviewed-on: https://go-review.googlesource.com/c/go/+/729023
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Michael Pratt <mpratt@google.com>

0 files changed, 0 insertions, 0 deletions