RSDK-10643: condition var refresh thread

[lia-viam]

so I think this is how we would want to do a condition_variable-based refresh thread in RobotClient, opening as WIP now for feedback though

[acmorrow]

If should_refresh_ is changing state, it should do so under the lock I think.

[lia-viam]

I think it is not changing state-- should_refresh_ is only mutated in this method and in the named constructor which sets it in the first place, so there may be a concurrent read happening here but there's no way for it to be read while being mutated

[acmorrow]

I'll argue that this is one of those cases where even if it is safe in practice, it is still better to do the check under the lock. It'll be annoying if someday we stand up TSAN and it whines about a leak here, and this isn't a hot path, so trying to eliminate the lock doesn't buy much of anything.

[acmorrow]

I mean race of course, not leak, since I'm talking about TSAN not ASAN.

[lia-viam]

done!

[acmorrow]

I'd even move the read of should_referesh_ under the lock, or even eliminate the check entirely. It should be fine in close to just unconditionally set should_refresh_ = false (under the lock), and then notify. If there is no refresh thread, it is just a big no-op.

[acmorrow]

Could this be while (should_refresh_)?

[acmorrow]

Might be possible to eliminate the bool by making refresh_interval_ an optional.

[lia-viam]

i think you're right, we're storing a T data; bool valid; which is basically just what an optional is under the hood.

(we could just have refresh_interval_ alone do all the work, but it's a little less self-documenting, although there are compact_optional types that allow this sort of thing)

[acmorrow]

I think it is probably worth doing. There's just less state to manage, which means less risk of things decohering. It'd also simplify RobotClient::with_channel:

From

    robot->refresh_interval_ = std::chrono::seconds{options.refresh_interval()};
    robot->should_refresh_ = (robot->refresh_interval_ > std::chrono::seconds{0});
    if (robot->should_refresh_) {
        robot->refresh_thread_ = std::thread{&RobotClient::refresh_every, robot.get()};
    }

To

    if (options.refresh_interval() != 0) {
        robot->refresh_interval_.reset(options.refresh_interval());
        robot->refresh_thread_ = std::thread{&RobotClient::refresh_every, robot.get()}
    }

[acmorrow]

condvars can (in theory) have spurious wakeups (the reasons are somewhat historically [interesting] (https://stackoverflow.com/a/8594964)). Potentially, this could use wait_until to reenter the wait state should that happen. However, I'm not sure the complexity is warranted. Is there a risk to doing an early refresh in the event of a spurious wakeup?

[lia-viam]

this one is interesting and relates to my answer about your other comments regarding the if and while conditions. spurious wakeups are the reason condvar code often needs an extra flag variable--if there's a wakeup we can tell if it's spurious or not by checking the value of bool no_i_really_meant_to_wake_you_up. all of the condition_variable::wait[_preposition] functions have an overload taking a predicate argument, and per cppreference (and, eg, libcxx) this is equivalent to while(!pred) wait(lk).

if we don't mind doing an extra refresh during a spurious wakeup, which I agree is probably harmless, it occurred to me that this whole function could just be

void RobotClient::refresh_every()
{
    std::unique_lock<std::mutex> lk{refresh_lock_}
    refresh_cv_.wait_for(lk, refresh_interval_, [this]{
        if (should_refresh_) refresh();
        return !should_refresh_;
    });
}

regarding the optional suggestion below though this would make me feel a bit weird about turning refresh_interval_ into an optional because I think that gets a little harder to reason about the validity of reading from refresh_interval_ or having to handle it being a nullopt

[acmorrow]

Maybe can be part of the loop condition?

[acmorrow]

Can refresh fail for legit but transient reasons like network being down? If so, is it actually correct to permanently shut down the refresh loop?

[lia-viam]

hmm good question, I was just copying the semantics from the previous implementation and adding a log statement. I imagine distinguishing types of errors could be done by catching GRPException separately and inspecting it

[acmorrow]

I'll argue that this is one of those cases where even if it is safe in practice, it is still better to do the check under the lock. It'll be annoying if someday we stand up TSAN and it whines about a leak here, and this isn't a hot path, so trying to eliminate the lock doesn't buy much of anything.

[acmorrow]

I think this is going to terminate after one timeout, but tbh I always get this stuff wrong, so don't let me mislead you if you have thought it through and this actually does what is intended.

[lia-viam]

you were right, classic case of a solution that is simple, elegant, and wrong 🙂

[acmorrow]

I'm generally skeptical of the predicate taking condition things because they are so easy to get wrong, and require too much thinking.

Here is how I think I would write this loop unrolled:

std::unique_lock<std::mutex> lk{refresh_lock_};
// Terminate if we ever see `should_refresh` as `false` while holding the lock
while (should_refresh_) {
    // Compute the absolute time on which the next refresh should happen
    const auto wake_time = std::chrono::steady_clock::now + refresh_interval_;
    // Loop until deadline is reached or `should_refresh_` becomes `false`
    while (should_refresh_) {
        const auto wait_result = refresh_cv.wait_until(lk, wake_time);
        if (wait_result == cv_status::timeout && should_refresh_) {
            // We hit the deadline: do a refresh
            try {
                refresh();
            } catch (...) { ... }
            // Break out of inner loop to compute next deadline
            break;
        } else {
            // If this was a spurious wake, then `should_refresh_` is still `true`, and
            // `wait_until` will be reinvoked with existing deadline. If this was not
            // a spurious wake, then we were notified, and both the inner and outer loop
            // will terminate.
        }
    }
}

But it is possible that the predicate version is more or less equivalent to that, in which case it should be fine.