Pi Zero W: intermittent loss of approximately 4295 seconds with 6.12 ARMv6 kernels

未關閉
#7,764 2 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
5/5
預估耗時
一週以上
新手友好度
35/100
Issue 類型
缺陷
描述清晰度
基本清楚
活躍度
活躍
技術堆疊
c, linux

研究方向

Start by examining the BCM2835 clocksource and clockevent handling for the 32-bit 1 MHz timer described in the issue. The report does not identify a failing kernel file or provide a deterministic reproduction, so first gather low-overhead diagnostics that distinguish clock accounting errors, clock-set operations, and prolonged stalls. Done means establishing evidence for the source of the roughly 4295-second clock deficit.

由索引模型根據 Issue 內容生成。

描述

Describe the bug

Two Raspberry Pi Zero W Rev 1.1 boards running 32-bit Raspbian Bookworm have intermittently fallen behind by about 71 min 35 s during operation. The application detects a large NTP offset and corrects the clock forward. The trigger is unknown.

Board Running kernel Observation
A 6.12.96+rpt-rpi-v6 2026-09-27: NTPsec recorded +4294.936 s and exited; the journal confirms a subsequent clock correction.
A 6.12.109+rpt-rpi-v6 2026-10-13: +4294.035920 s measured and corrected.
B 6.12.96+rpt-rpi-v6 2026-10-19: application logged a 4294 s correction; no journal confirmation of the step.
A 6.12.109+rpt-rpi-v6 2026-11-17: +4294.844522 s measured; residual after correction −0.011612 s.

On an earlier incident on board A, two public NTP references and the LAN reference independently agreed that the local clock was behind by about 4296 s.

Steps to reproduce the behaviour

No deterministic reproduction. The boards run headless, with Wi-Fi and Bluetooth active, and an application periodically comparing the local clock against NTP. Expected: the clock advances correctly during unattended operation. Observed: a sporadic large deficit, subsequently corrected forward. No stock-image or kernel-only A/B test has been performed.

Device (s)

Raspberry Pi Zero W / WH

System

Raspberry Pi Zero W Rev 1.1; Raspbian GNU/Linux 12 (Bookworm), armv6l.

Board A on 2026-11-17: 6.12.109+rpt-rpi-v6, package 1:6.12.109-1+rpt1 (build 2026-09-11); boot firmware 2026-09-14 (a86983925695); get_throttled=0x0. The refreshed Bookworm ARMv6 candidate 1:6.12.115-1+rpt1 has not yet been tested in this setup.

All four comparison boards reported active clocksource timer, mask 0xffffffff, and a 32-bit 1 MHz timer in the boot journal. All loaded the same BCM43430 WLAN firmware (7.45.98, FWID 01-8e14b897) and used NTPsec 1.2.2.

Logs

2026-11-17, monotonic journal timeline (seconds since boot):

  • 313440.978: locally generated Wi-Fi disconnection; connection to a different AP 0.567 s later. DHCP renewed the same IPv4 address about 1.0 s after the disconnection.
  • 313613.414: NTPsec stopped normally for the application-initiated correction, about 172 s after disconnection. No panic exit in this window.
  • 313619.298: Tailscale detected the forward clock step during correction; this is not evidence of suspend.
  • 313620.930: NTPsec started again.
  • 313621.682: correction verified at 08:04:45 CET, applied correction +4294.844522 s, residual −0.011612 s.

About 45 minutes after correction two public NTP sources reported offsets below 2 ms. The latest large offset itself was measured against the LAN reference. The October journal also showed a nearby Wi-Fi reassociation but does not establish a causal link. No matching kernel error was found in the available journal.

Additional context

Two comparable Zero W Rev 1.1 boards still running 6.1.21+ with older boot firmware have no known matching incidents. Kernel, boot firmware and some user-space packages changed together, so this is an association, not an isolated kernel regression.

A full 32-bit timer period at 1 MHz is 4294.967296 s. The latest correction was about 0.123 s smaller. Could this signature indicate lost time in BCM2835 clocksource or clockevent handling? What low-overhead diagnostics would help distinguish a clock accounting error, an unexpected clock-set operation, and a prolonged stall?

A historical FreeBSD report gives a similar offset but does not establish a Linux cause: https://lists.freebsd.org/pipermail/freebsd-arm/2012-December/004443.html

主要語言
C
星號
13.3k
分支
5.5k
平均合併
4 天 10 小時
30 天內合併 PR
31

環境準備

這個專案沒有提供開發容器、Dockerfile 或貢獻指南,環境需要你自己搭建:先看它的 README,通用步驟見我們的新手貢獻指南。

從這裡開始

  1. 先讀完整個 Issue,再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫,在一個分支上完成修改。
  4. 送出 Pull Request,並在描述裡引用這個 Issue 編號。

raspberrypi/linux 的其他 Issue

查看 raspberrypi/linux 的全部 Issue

相似的 Issue

更多 C Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。