[codethinki]

A strand for cth_coro: a serial queue on the workers of a scheduler. jvk needs one per Vulkan queue, so submits of one queue never run at the same time, without a thread per queue.

context

• New abstract cth::co::context: post(), await(handle), await(time point), owns_thread() and current(). Protected non-virtual destructor.
• scheduler and strand implement it, both final.
• executor, the payload, the awaiters and the this_coro tag take context const& instead of scheduler const&.
• Renames: scheduler_payload -> context_payload, payload.scheduler() -> payload.context(), rebind_scheduler() -> rebind_context(), this_coro::scheduler -> this_coro::context, executor::scheduler() -> executor::context().

strand

• explicit strand(scheduler const&): its jobs run one at a time and in post order, on the workers of the scheduler. The scheduler outlives the strand.
• await callbacks and timer callbacks run inside the strand.
• A coroutine holds the strand between two suspension points, not across a co_await.

context::current()

• The context of the job on the calling thread: the strand inside a strand job, the scheduler in any other worker job, nullptr elsewhere.
• schedule_awaiter and the spawn final awaiter decide with it. Before, they used owns_thread(), which is also true inside a strand job of the same scheduler, so a coroutine could not leave a strand and a spawn caller could end up inside one.

Shared asio code exists once: src/utility/asio/ops.hpp (post_on, await_on), the macro in asio/macro.hpp.

Tests: 1000 jobs on 4 workers run one at a time and in order, awaits run inside the strand, executor{strand}.schedule(), a coroutine leaves a strand, spawn from and into a strand, current(), a scheduler job runs while a strand job runs. 297/297 on gcc and clang, debug and release. ASan: 0 leaks.

[codethinki]

Replaced by #25.