[mcollina]

This is the undici counterpart of nodejs/node#67292.

Problem

The keep-alive timer of an idle HTTP/1.1 socket cannot fire while the event loop is blocked. Before reusing an idle socket, undici yields one setImmediate (scheduleIdleSocketValidation, added for GHSA-35p6-xmwp-9g52), so that a FIN or RST that has already been received is processed first. That only works if a poll phase runs between the dispatch and the check phase.

With a server in a separate process (keepAliveTimeout = 5000), one request, then a 6s synchronous block, then a second request:

Second request dispatched from	Processed before the validation completes	Result on main
poll phase (I/O callback, server request handler, await after I/O)	nothing: the check phase comes next	ECONNRESET: the dead socket is reused
timers phase	the poll phase reads the FIN	reconnects
check phase (setImmediate)	the keep-alive timer fires	reconnects

The poll-phase case is the common one, for example a proxy that does synchronous work in its request handler before calling upstream. A shorter version of the same problem: a block longer than the client's keep-alive timeout but shorter than the server's uses up the keepAliveTimeoutThreshold margin, and the request is sent just before the server closes the connection.

Fix

• When resumeH1 arms the keep-alive timer, record performance.now() on the socket.
• Before reusing an idle socket, destroy it with InformationalError('socket idle timeout') if its keep-alive timeout has passed. Nothing is running on the socket, so the pending requests are sent on a new connection, the same as when the timer fires.

The time is recorded when the timer is armed, so the check never expires a socket before the timer would have. It uses the global performance.now(), so it follows @sinonjs/fake-timers together with setTimeout. The cost is one performance.now() call when a socket becomes idle and one when it is reused.

The check does not depend on kSocketUsed, so it also covers the experimental Milo parser. parser-h1.js never sets that symbol, which is private to client-h1.js. As a result, the existing idle socket validation does not run with Milo. That is a separate issue and is left for a follow-up.

Not covered

As in the Node.js change:

• a server close that is still in flight when the request is sent (the general keep-alive race; the retry interceptor handles this for idempotent requests)
• time spent blocked before the socket becomes idle

HTTP/2 is not changed.

Testing

• test/client-keep-alive-blocked-loop.js blocks inside an fs callback (poll phase) and checks that an expired socket is replaced and that a socket that has not expired is reused. The expiry case fails on main and passes with this change, with both llhttp and UNDICI_USE_MILO=1.
• The out-of-process reproduction above passes for all three phases, and a 1s block still reuses the socket.
• test/*.js: 1664 passed, 0 failed, 3 skipped. test/node-test/**: 245 passed, 0 failed, 5 skipped.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Sfj4nEWQ4ovycNDjbvXcRs

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.68%. Comparing base (76f0b3e) to head (8f1768a).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #6171   +/-   ##
=======================================
  Coverage   91.67%   91.68%           
=======================================
  Files         113      113           
  Lines       44419    44467   +48     
=======================================
+ Hits        40721    40768   +47     
- Misses       3698     3699    +1     

☔ 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.