server: fix data races found by the nightly ThreadSanitizer job

[Frauschi]

What & why

The nightly Sanitizers workflow's ThreadSanitizer job has been failing on every threaded roundtrip since the TSan build actually started running (the earlier -Werror fix let it get far enough to execute the tests). Root cause: two genuine data races.

1. stopping flag — src/server.c

WolfCertServer.stopping was a plain volatile int, written by wolfcert_server_stop() (from another thread in the test harness, or from the wolfcert-server CLI's SIGINT/SIGTERM handler) and polled by the accept loop in wolfcert_server_run(). volatile establishes no happens-before, so TSan flags it. Separately, stop() close()d listen_fd while the accept loop still read it — a second race hidden behind the first.

2. port handoff — tests/integration/test_tls_http.c

srv_thread published the ephemeral listener port into a main-thread stack struct that main spin-polled without synchronization.

The fix

• stopping → wolfSSL_Atomic_Int (wc_port.h), accessed via WOLFSSL_ATOMIC_LOAD / WOLFSSL_ATOMIC_STORE. Gives the cross-thread handoff a happens-before edge and an async-signal-safe store for the CLI.
• The accept loop now poll()s the listener on a short timeout and re-checks the flag itself. stop() only sets the flag and never touches listen_fd (closed by free() after the serving thread is joined). This removes the second race and drops the reliance on shutdown()/close() waking a blocked accept(), which is not portable to BSD/macOS.
• test_tls_http's port field → wolfSSL_Atomic_Int as well.
• Hardening: the accept loop retries transient per-connection errors (ECONNABORTED / ECONNRESET) instead of tearing the listener down over a single misbehaving peer.

Embedded / portability

wc_port.h's atomics degrade to volatile int under WOLFSSL_NO_ATOMICS, so the client/embedded path is unchanged — on a no-atomics target the field is literally volatile int again, as before. The change is confined to the server subsystem (WOLFCERT_ENABLE_SERVER) and one integration test.

Verification

• Local: 12/12 roundtrip|tls_http pass clean under -fsanitize=thread (halt_on_error=1); full 24-test suite green.
• CI: Sanitizers workflow dispatched on this branch — ThreadSanitizer, ASan+UBSan, and valgrind all ✅ (run 29482351976).

[wolfSSL-Fenrir-bot]

Fenrir Automated Review — PR #9

Scan targets checked: wolfcert-bugs, wolfcert-src

No new issues found in the changed files. ✅

[Copilot]

Pull request overview

Fixes ThreadSanitizer-reported data races in the server shutdown path and in an integration test’s cross-thread port publication, improving correctness of threaded operation under TSan and real concurrency.

Changes:

• Replace WolfCertServer.stopping with wolfSSL_Atomic_Int and use WOLFSSL_ATOMIC_LOAD/STORE for cross-thread stop signaling.
• Make the accept loop poll() with a short timeout instead of blocking in accept(), and stop closing listen_fd from wolfcert_server_stop().
• Fix the test_tls_http port handoff race by publishing/polling the port via wolfSSL_Atomic_Int.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
tests/integration/test_tls_http.c	Uses wolfSSL_Atomic_Int for the server port handoff between threads to satisfy TSan.
src/server.c	Reworks the accept loop to poll and uses atomic stop signaling; removes cross-thread listen_fd shutdown/close.
src/internal.h	Updates WolfCertServer to store the stop flag as wolfSSL_Atomic_Int.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[philljj]

Looks good, just one nit on poll(2) event handling.