bluerobotics / bluerobotics/ping-cpp
Reading from a UDP link busy-loops
- 主要言語
- C++
- スター
- 19
- フォーク
- 17
- PR マージ指標
- 30日以内にマージされた PR はありません
説明
It would appear the API provides no way to structure read handling so it does not 100% saturate a CPU core while waiting for received messages. This was maybe tolerable when all messages were synchronous command responses, but with the new auto transmit function, my driver needs to just sit and wait for asynchronous receive data all of the time.
To start with, the provided test-device-ping360.cpp example uses `PingDevice::waitMessage()` to wait for auto data packets. Internally, `waitMessage()` calls `PingDevice::read()` in a tight loop until its timeout expires. There is a comment here that says `read()` blocks up to 0.1s, but this comment is a lie (at least for UDP, I didn't check the serial path). Inside `PingDevice::read()`, we just call the port `read()` method exactly once, and then attempt to parse any data returned from it. `UdpLink::read()` always returns immediately with whatever amount of data it was able to pull out of its buffer at that moment, so there is no blocking happening.
Even if the test case using `waitMessage()` didn't have this problem, I would ideally like to structure my driver such that I'm not continuously polling for new data. I tried setting things up with `AbstractLink::doOnReceived()` to let me know when data is available, so I could read until the available data was exhausted and handle any packets that were parsed in the process, but it turns out this is impossible with the given API. Since `PingDevice::read()` reads exactly one byte and returns null in the case that *either* nothing was read *or* a byte was read but it didn't make a message yet (notwithstanding the API documentation on `read()` claiming that it reads until no data is left in the buffer), I have to call `read()` in a loop until the buffer is empty, but I have no good way of knowing that. I tried calling `read()` exactly the number of times as the number of entries in the vector given to me by the `onReceived` signal, but it turns out this doesn't work either, because `UdpLink` fires the signal *before* it puts the data in its buffer, meaning I can get notified 30 bytes were received, and call `read()` 30 times, getting nothing each time, *before* `UdpLink` makes its newly received data available to be read.
To summarize:
* `PingDevice::waitMessage()` with UDP busy-loops for the duration of its timeout
* `UdpLink::read()` does not block the way the comment in `PingDevice::waitMessage()` claims it does
* `PingDevice::read()` does not exhaust the port buffer the way its API comment claims it does
* Ideally, `UdpLink` would emit the `onReceived` signal only after the data it's signalling about is actually available
コントリビューションガイド
このリポジトリのコントリビューションガイドは索引されていません
調査の方向性
test-device-ping360.cpp と、issue で説明されているエントリポイント PingDevice::waitMessage()、PingDevice::read()、UdpLink::read()、AbstractLink::doOnReceived() から始めます。UDP データの到着、バッファリング、シグナルの順序を追跡し、その後、待機がビジーループにならないこと、読み取りがドキュメント化された動作に従うこと、データが利用可能になった後に onReceived が発火することを確認します。
索引モデルが issue の本文から書いたものです。
評価
- 技術スタック
- cpp
- 領域
- networking
- issue の種類
- バグ
- 難易度
- 4/5
- 見積もり時間
- 3〜5日
- 活発さ
- 停滞
- 明瞭さ
- おおむね明確
- 初心者へのやさしさ
- 42/100