coroutine: exception: workaround broken destroy coroutine handle in await_suspend

gcc 15.2 and above crash if a coroutine handle is destroyed when
await_suspend for that coroutine is called [1]. An internal reference
counter is accessed after the coroutine is destroyed.

Work around this by scheduling the work to a task. This is slower,
but doesn't crash, and exceptions are supposed to be somewhat rare.

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121961

Author: avikivity

include/seastar/core/coroutine.hh

@@ -36,6 +36,20 @@ namespace seastar {
 
 namespace internal {
 
+
+inline
+void
+execute_involving_handle_destruction_in_await_suspend(std::invocable<> auto&& func) noexcept {
+#if defined(__GNUC__) && !defined(__clang__) && (__GNUC__ > 15 || (__GNUC__ == 15 && __GNUC_MINOR__ >= 2))
+    // See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121961
+    memory::scoped_critical_alloc_section _;
+    schedule(new lambda_task(std::forward<decltype(func)>(func)));
+#else
+    std::invoke(std::forward<decltype(func)>(func));
+#endif
+}
+
+
 template <typename T = void>
 class coroutine_traits_base {
 public:

include/seastar/coroutine/exception.hh

@@ -43,8 +43,10 @@ struct exception_awaiter {
 
     template<typename U>
     void await_suspend(std::coroutine_handle<U> hndl) noexcept {
+      execute_involving_handle_destruction_in_await_suspend([hndl, eptr = std::move(eptr)] () mutable {
         hndl.promise().set_exception(std::move(eptr));
         hndl.destroy();
+      });
     }
 
     void await_resume() noexcept {}