[quinnj]

Every HTTP/2 frame the client or server wrote went through a fresh IOBuffer: _serialize_frame copied the payload, the 9-byte header went into its own vector, both were written into the IOBuffer, and take! produced the bytes to send. Returning flow-control credit wrote two WINDOW_UPDATE frames (connection and stream) as two separately locked writes, after every body_read!. Reading allocated a 9-byte vector for each frame header and a 4-byte payload for each WINDOW_UPDATE.

Now:

• A frame is serialized straight into one exact-size vector (_frame_bytes), without the payload copy or IOBuffer, and write_frame! writes it in one call.
• A credit's connection and stream WINDOW_UPDATE frames are built in one 26-byte vector and sent in one write under one lock (_window_update_frames), on both client and server.
• When the connection reader already buffers them, a frame header and a WINDOW_UPDATE frame are parsed in place.
• _header_block_frames no longer copies a header block that fits in one frame; every caller writes the frame immediately.

The bytes on the wire are unchanged, in the same order.

Whole-process allocations for a 1 MiB HTTP/2 upload (client and server in one process, --threads=4,1, minimum of 5):

Julia	master	this PR
1.10	1899	901
1.11	2633	1157
1.12	2811	1162
1.13	2532	1156

Julia 1.12, in-process round trips (latency: median and min–max of per-run medians over 4 interleaved runs per build, heavily loaded machine):

	master	this PR
1 MiB POST, allocations per request	2884	1258
1 MiB POST, latency	24.1 ms (22.8–26.1)	23.3 ms (18.3–25.6)
GET, allocations per request	449	410
GET, latency	443 µs (340–495)	433 µs (372–480)
socket writes per credit returned	2	1

Reading a buffered WINDOW_UPDATE frame allocates 16 bytes (the returned frame) instead of 144–160 on Julia 1.10 through 1.13.

Verification

• New test in test/http2_frame_tests.jl: _window_update_frames produces the same bytes as write_frame! of the two frames; frames read through a 4-byte _ConnReader buffer (headers and payloads split across reads) and through a prefilled one parse the same; reading a buffered WINDOW_UPDATE allocates at most 64 bytes (144–160 on master).
• http2_frame_tests.jl, http2_server_tests.jl, http2_client_tests.jl, http2_settings_tests.jl, http2_refused_stream_retry_tests.jl pass on 1.12. Rebased on Fix HTTP/2 server receive-window accounting #1452, whose receive-window accounting in _send_h2_server_window_updates! is kept. The full suite with this and my other open allocation changes combined passes on 1.12 except trim_compile_tests.jl, which fails the same way on master locally (a verifier warning inside Reseau.TLS._read_some!).

Review

• _serialize_frame has one caller, _frame_bytes, which copies the payload into the new vector, so returning the frame's own payload vector is safe.
• The in-place paths run only for a _ConnReader with the whole header (9 bytes) or WINDOW_UPDATE payload (4 bytes) buffered; otherwise the generic path runs. The fast WINDOW_UPDATE path applies the same length, reserved-bit mask, and zero-increment checks and errors as the generic one.
• Window updates were previously two lock acquisitions, so another writer's frame could fall between them; now they are adjacent. Both orders are valid HTTP/2. Errors are handled as before (EOF and closing errors are ignored; others rethrown), and the client still sets its write deadline before writing.
• The client serializes a frame before taking the write lock rather than inside it; frames are immutable and serialization touches no shared state.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.61%. Comparing base (84a7fe5) to head (33afa85).
⚠️ Report is 29 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1487      +/-   ##
==========================================
+ Coverage   91.55%   91.61%   +0.05%     
==========================================
  Files          31       31              
  Lines       12972    13081     +109     
==========================================
+ Hits        11877    11984     +107     
- Misses       1095     1097       +2     

☔ 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.