[binaryfire]

Swoole\Coroutine\Http2\Client can only read a header block that arrives in a single, unpadded HEADERS frame, and both the client and the server keep the padding of DATA frames in the body.

Client

• A server has to split a header block larger than the client's 16 KB max frame size into HEADERS and CONTINUATION frames (RFC 9113 §4.3), and nginx does this for a response with many large headers. The client ignores CONTINUATION frames and decodes every HEADERS frame as a complete block, so recv() fails with -523 (HPACK error).
• With the PADDED flag, the pad length and padding of a HEADERS frame go to the HPACK decoder, so responses from a peer that pads (such as nghttpd -b 255) fail the same way. In DATA frames they end up in the response body.
• With the PRIORITY flag, the client skips 5 bytes without checking the frame length. A shorter HEADERS frame makes the decoder read past the end of the frame (Valgrind reports invalid reads).

Header block fragments are now decoded as they arrive, with END_HEADERS marking the last one, the same way the server does it. The client waits for END_HEADERS before it finishes a header block, and keeps END_STREAM from the HEADERS frame that started it. Padding is removed from HEADERS and DATA frames. DATA padding still counts toward flow control. The connection is closed with PROTOCOL_ERROR or FRAME_SIZE_ERROR on:

• a frame other than a CONTINUATION on the same stream inside an unfinished header block
• a CONTINUATION frame outside one
• padding that doesn't fit in the frame
• a HEADERS frame too short for its priority fields

Server

The server already removes the padding from HEADERS frames but not from DATA frames, so when a client pads DATA (such as nghttp -b 255) the padding ends up in $request->rawContent(). DATA is now handled the same way as HEADERS. #6389 also changes this DATA case, so whichever lands second needs a small rebase.

Tests

continuation_frames.phpt and padded_frames.phpt drive the client from a raw socket peer, including the error cases and a WINDOW_UPDATE check that padding still counts. padded_data.phpt covers the server in BASE, PROCESS and coroutine mode. All three fail on master. I also checked the client against nginx (large headers) and nghttpd -b 255 and the server against nghttp -b 255, and ran the new tests under Valgrind and on a ZTS thread build, plus a THREAD mode server probe.

Hypervel's HTTP/2 and gRPC clients use Coroutine\Http2\Client, and this broke its CONTINUATION regression test.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.03%. Comparing base (9cd45e0) to head (89651ea).

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #6399      +/-   ##
==========================================
- Coverage   84.08%   84.03%   -0.05%     
==========================================
  Files         115      115              
  Lines       20237    20237              
  Branches     3637     3637              
==========================================
- Hits        17016    17007       -9     
- Misses       3221     3230       +9     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.