embassy/embassy-net-driver-channel
2023-12-04 13:15:10 +00:00
..
src net/driver: remove Medium, make HardwareAddress non_exhaustive. 2023-10-18 05:28:16 +02:00
Cargo.toml update release version in examples and other crates 2023-12-04 13:15:10 +00:00
CHANGELOG.md net/driver: remove Medium, make HardwareAddress non_exhaustive. 2023-10-18 05:28:16 +02:00
README.md net/driver: remove Medium, make HardwareAddress non_exhaustive. 2023-10-18 05:28:16 +02:00

embassy-net-driver-channel

This crate provides a toolkit for implementing embassy-net drivers in a higher level way than implementing the embassy-net-driver trait directly.

The embassy-net-driver trait is polling-based. To implement it, you must write the packet receive/transmit state machines by hand, and hook up the Wakers provided by embassy-net to the right interrupt handlers so that embassy-net knows when to poll your driver again to make more progress.

With embassy-net-driver-channel you get a "channel-like" interface instead, where you can send/receive packets to/from embassy-net. The intended usage is to spawn a "driver task" in the background that does this, passing packets between the hardware and the channel.

A note about deadlocks

When implementing a driver using this crate, it might be tempting to write it in the most straightforward way:

loop {
    // Wait for either..
    match select(
        // ... the chip signaling an interrupt, indicating a packet is available to receive, or
        irq_pin.wait_for_low(),
        // ... a TX buffer becoming available, i.e. embassy-net wants to send a packet
        tx_chan.tx_buf(),
    ).await {
        Either::First(_) => {
            // a packet is ready to be received!
            let buf = rx_chan.rx_buf().await; // allocate a rx buf from the packet queue
            let n = receive_packet_over_spi(buf).await;
            rx_chan.rx_done(n);
        }
        Either::Second(buf) => {
            // a packet is ready to be sent!
            send_packet_over_spi(buf).await;
            tx_chan.tx_done();
        }
    }
}

However, this code has a latent deadlock bug. The symptom is it can hang at rx_chan.rx_buf().await under load.

The reason is that, under load, both the TX and RX queues can get full at the same time. When this happens, the embassy-net task stalls trying to send because the TX queue is full, therefore it stops processing packets in the RX queue. Your driver task also stalls because the RX queue is full, therefore it stops processing packets in the TX queue.

The fix is to make sure to always service the TX queue while you're waiting for space to become available in the RX queue. For example, select on either "tx_chan.tx_buf() available" or "INT is low AND rx_chan.rx_buf() available":

loop {
    // Wait for either..
    match select(
        async {
            // ... the chip signaling an interrupt, indicating a packet is available to receive
            irq_pin.wait_for_low().await;
            // *AND* the buffer is ready...
            rx_chan.rx_buf().await
        },
        // ... or a TX buffer becoming available, i.e. embassy-net wants to send a packet
        tx_chan.tx_buf(),
    ).await {
        Either::First(buf) => {
            // a packet is ready to be received!
            let n = receive_packet_over_spi(buf).await;
            rx_chan.rx_done(n);
        }
        Either::Second(buf) => {
            // a packet is ready to be sent!
            send_packet_over_spi(buf).await;
            tx_chan.tx_done();
        }
    }
}

Examples

These embassy-net drivers are implemented using this crate. You can look at them for inspiration.

Interoperability

This crate can run on any executor.

License

This work is licensed under either of

at your option.