After my first attempt at fixing the client disconnects
(https://github.com/OttoWinter/AsyncTCP/pull/4) got merged
into AsyncTCP, it turned out that there was regression
for some devices: the connection stability actually went
down instead of up.
After a lot of debugging and discussion with @glmnet (some of
the results can be found in the above pull request discussion),
we came up with an improved fix for the disconnect issues.
**Changed:**
The code that checks for ACK timeouts has been simplified in
such way, that only two timestamps are now used to determine if
an ACK timeout has happened: the time of the last sent packet
(this was already recorded), and the time of the last received
ACK from the client (this has been added).
Using these timestamps, there is no more need for a separate field
to keep track if we are waiting for an ACK or not (`_pcb_busy`).
Therefore, this field was completely removed from the code.
While I was at it, I renamed a few variables to make the code
easier to read and more consistent.
**Results:**
I connected Home Assistant plus 8 OTA loggers at the same time,
using very verbose logging output. This normally was an easy way
to trigger the disconnect errors.
It turned out, this solution runs as solid for me, as when disabling
the ACK timeout checks completely
(using `AsyncClient::setAckTimeout(0)`).
The AsyncClient::send() methods sets a boolean to true after pushing
data over the TCP socket successfully using tcp_output(). It also sets
a timestamp to remember at what time the data was sent.
The AsyncClient::_sent() callback method reacts to ACKs coming from
the connected client. This method sets the boolean to false.
In the AsyncClient::_poll() method, a check is done to see if the
boolean is true ("I'm waiting for an ACK") and if the time at which
the data was sent is too long ago (5000 ms). If this is the case,
a connection issue with the connected client is assumed and the
connection is forcibly closed by the server.
The race condition is when these operations get mixed up, because
of multithreading behavior. The _sent() method can be called during
the execution of the send() method:
1. send() sends out data using tcp_output()
2. _sent() is called because an ACK is processed, sets boolean to false
3. send() continues and sets boolean to true + timestamp to "now"
After this, the data exchange with the client was successful. Data were
sent and the ACK was seen.
However, the boolean ended up as true, making the _poll() method think
that an ACK is still to be expected. As a result, 5000 ms later, the
connection is dropped.
This commit fixes the code by first registering that an ACK is
expected, before calling tcp_output(). This way, there is no race
condition when the ACK is processed right after that call.
Additionally, I changed the boolean to an integer counter value.
The server might send multiple messages to the client, resulting in
multiple expected ACKs. A boolean does not cover this situation.
Co-authored-by: Maurice Makaay <mmakaay1@xs4all.net>
Mathieu Carbou
Jesse Hills
Jimmy Hedman
Kuba Szczodrzyński
Guillermo Ruffino
Maurice Makaay
Otto Winter
Maurice MakaayandMaurice Makaay