TCP connections fail after 49 days due to 32-bit counter overflow in XNU kernel, affecting Mac servers and CI systems
Photon discovered this networking nightmare while running their iMessage monitoring service. Their Mac fleet started mysteriously refusing new TCP connections after exactly 49 days, 17 hours, 2 minutes, and 47.296 seconds of continuous uptime. Existing connections worked fine. Ping worked perfectly. New outbound connections? Dead in the water.
The culprit lives deep in macOS’s XNU kernel: a 32-bit counter called tcp_now that tracks milliseconds since boot. When this counter overflows after 2^32 milliseconds, Apple’s code freezes like a deer in headlights. The TIME_WAIT garbage collection system—responsible for cleaning up closed connections—stops working entirely.
No kernel panic, no clear logs—just accumulating network chaos.
Photon’s live reproduction tells the whole story. Before overflow, their systems maintained a healthy ~200 TIME_WAIT connections during normal load. After the magical 49.7-day mark, these connections started piling up like unopened mail:
SYN_SENT connections climbed past 3,000. System load spiked to 49.74.
This explains those mysterious Apple Community forum posts about TCP failures after weeks of uptime, and why Podman users reported networking stalls on macOS 12 VMs. The symptoms were always there—someone just needed to connect the dots to that precise timestamp.
Consumer Macs dodge this bullet, but your server infrastructure needs attention.
Your MacBook Pro probably updates and reboots frequently enough to never see this bug. But that Mac Pro running CI/CD builds in your data center? That Mac mini cluster handling automated workflows? They’re sitting ducks.
The only current fix is rebooting before hitting 49.7 days—hardly elegant for enterprise deployments. Photon is developing a non-reboot workaround, but Apple hasn’t responded yet to what amounts to a fundamental violation of RFC 7323’s timestamp handling expectations.
Check your uptime with uptime and schedule those maintenance windows accordingly. In the world of always-on infrastructure, even Apple’s rock-solid Unix foundation has an expiration date—and it’s precisely 49 days, 17 hours, 2 minutes, and 47.296 seconds away.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。