On Tue, 1 Sep 2015, Nuno Gonçalves wrote: > There is a regression on the clock system since v3.16-rc5-111-g4396e05 > [1], > [1] http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=4396e058c52e167729729cf64ea3dfa229637086 That commit has absolutely nothing to do with NTP. I fear your bisect went down the wrong road somewhere. > where the clock doesn't apply frequency offsets above about > 1000ppm [2]. This looks pretty familiar. The issue was introduced with commit 5e5aeb4367b (time: adjtimex: Validate the ADJ_FREQUENCY values). That patch was tagged for stable, so it got backported. The fix is in commit 29183a70b0b82 (ntp: Fixup adjtimex freq validation on 32-bit systems). That commit was tagged for stable as well, but with the extra '#3.19+' limitation. So in the worst case 5e5aeb4367b hit a stable tree < 3.19, but 29183a70b0b82 did not. Thanks, tglx