This PR adds a LargeResponse example that we did not have in the project to show the usage of a custom response and the AwsResponseFiller callback.

This example was also built in order to serve as a MRE for these 2 issues:

• Lost writes using AsyncCallbackResponse during low memory #242
• AsyncAbstractResponse truncates 85 octets in the non-empty segment #315

Endpoint using AwsResponseFiller

curl -v http://192.168.4.1/1 | grep -o '.' | sort | uniq -c

5652 A
4308 B
5760 C
 280 D

ESP32:

Filling 'A' @ index: 0, maxLen: 5652, toSend: 5652
Filling 'B' @ index: 5652, maxLen: 4308, toSend: 4308
Filling 'C' @ index: 9960, maxLen: 5760, toSend: 5760
Filling 'D' @ index: 15720, maxLen: 280, toSend: 280

Endpoint with a custom response:

curl -v http://192.168.4.1/2 | grep -o '.' | sort | uniq -c

5675 A
4308 B
2888 C
3129 D

ESP32:

Filling 'A' @ sent: 0, buflen: 5675
Filling 'B' @ sent: 5675, buflen: 4308
Filling 'C' @ sent: 9983, buflen: 2888
Filling 'D' @ sent: 12871, buflen: 3129

So everything works fine in the example.
I am sending 16K, I receive 16K and I also receive the right number of characters.

@ESP32Async/devs : build pass. I am going to merge in main do that we can rebase and use these examples as a base to work on for the prs / fixes. Let me know if something is off otherwise i will merge. Thank you!

looks OK to me, but with a quick test on my side it also does not reproduce the issue - strange. Will play with it more tomorrow

For now i will merge - these are good examples we do not have. And besides they will be helpful for your tests ;-)

I cracked it.

So this is how it works:
That space that _tcp is writing to identified by CONFIG_TCP_SND_BUF_DEFAULT (and is value-matching with default TCP windows size which is very confusing itself).
The space returned by client()->write() and client->space() somehow might not be atomically/thread synced (had not dived that deep yet). So if first call to _fillBuffer is done via user-code thread and ended up with some small amount of data consumed and second one is done by _poll or _ack? returns full size again! This is where old code fails.

If you change your class this way it will fail 100%.

class CustomResponse : public AsyncAbstractResponse {
public:
  explicit CustomResponse() {
    _code = 200;
    _contentType = "text/plain";
    _sendContentLength = false;
    // add some useless headers
    addHeader("Clear-Site-Data", "Clears browsing data (e.g., cookies, storage, cache) associated with the requesting website.");
    addHeader("No-Vary-Search", "Specifies a set of rules that define how a URL's query parameters will affect cache matching. These rules dictate whether the same URL with different URL parameters should be saved as separate browser cache entries");
  }

  bool _sourceValid() const override {
    return true;
  }

  size_t _fillBuffer(uint8_t *buf, size_t buflen) override {
    if (fillChar == NULL){
      fillChar = 'A';
      return RESPONSE_TRY_AGAIN;
    }
    if (_sent == RESPONSE_TRY_AGAIN) {
      Serial.println("Simulating temporary unavailability of data...");
      _sent = 0;
      return RESPONSE_TRY_AGAIN;
    }
    size_t remaining = totalResponseSize - _sent;
    if (remaining == 0) {
      return 0;
    }
    if (buflen > remaining) {
      buflen = remaining;
    }
    Serial.printf("Filling '%c' @ sent: %u, buflen: %u\n", fillChar, _sent, buflen);
    std::fill_n(buf, buflen, static_cast<uint8_t>(fillChar));
    _sent += buflen;
    fillChar = (fillChar == 'Z') ? 'A' : fillChar + 1;
    return buflen;
  }

private:
  char fillChar = NULL; //'A';
  size_t _sent = 0;
};

@vortigont : thanks ! I will push an update to the example in the working branch of your PR.