cont.c (ec_switch): prevent delayed/missed trap interrupt race

timer-thread may set trap interrupt with rb_threadptr_check_signal
at any time independent of GVL.  This means timer-thread may set
the trap interrupt flag on the previous execution context; causing
the flag to be unnoticed until a future ec switch (or lost
completely if the ec is done).

Note: I avoid relying on th->interrupt_lock here and use
atomics because we won't be able to rely on it for proposed lazy
timer-thread [Misc #14937].

This regression affects Ruby 2.5 as it was introduced by moving
interrupt_flag to `ec' which is an unstable pointer.  Ruby <= 2.4
was unaffected because vm->main_thread->interrupt_flag never
changed.

[ruby-core:88119] [Bug #14939]

git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@64062 b2dd03c8-39d4-4d8f-98ff-823fe69b080e

1 files changed, 10 insertions, 0 deletions

diff --git a/cont.c b/cont.c
index 082b47dae2..1f82a7944e 100644
--- a/cont.c
+++ b/cont.c
@@ -262,7 +262,17 @@ static inline void
 ec_switch(rb_thread_t *th, rb_fiber_t *fib)
 {
     rb_execution_context_t *ec = &fib->cont.saved_ec;
+
     ruby_current_execution_context_ptr = th->ec = ec;
+
+    /*
+     * timer-thread may set trap interrupt on previous th->ec at any time;
+     * ensure we do not delay (or lose) the trap interrupt handling.
+     */
+    if (th->vm->main_thread == th && rb_signal_buff_size() > 0) {
+        RUBY_VM_SET_TRAP_INTERRUPT(ec);
+    }
+
     VM_ASSERT(ec->fiber_ptr->cont.self == 0 || ec->vm_stack != NULL);
 }