linux-kernel - Re: [PATCH] timekeeping: Prevent coarse clocks going backwards

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <87tt6mq8jz.ffs@tglx>
Date: Fri, 18 Apr 2025 08:37:36 +0200
From: Thomas Gleixner <tglx@...utronix.de>
To: John Stultz <jstultz@...gle.com>
Cc: Miroslav Lichvar <mlichvar@...hat.com>, LKML
 <linux-kernel@...r.kernel.org>, Stephen Boyd <sboyd@...nel.org>,
 Anna-Maria Behnsen <anna-maria@...utronix.de>, Frederic Weisbecker
 <frederic@...nel.org>, Shuah Khan <shuah@...nel.org>,
 linux-kselftest@...r.kernel.org, kernel-team@...roid.com, Lei Chen
 <lei.chen@...rtx.com>
Subject: Re: [PATCH] timekeeping: Prevent coarse clocks going backwards

On Thu, Apr 17 2025 at 17:46, John Stultz wrote:
> On Sat, Apr 5, 2025 at 2:40 PM Thomas Gleixner <tglx@...utronix.de> wrote:
>> @@ -1831,6 +1847,8 @@ void timekeeping_resume(void)
>>         /* Re-base the last cycle value */
>>         tks->tkr_mono.cycle_last = cycle_now;
>>         tks->tkr_raw.cycle_last  = cycle_now;
>> +       /* Reset the offset for the coarse time getters */
>> +       tks->coarse_nsec = 0;
>>
>>         tks->ntp_error = 0;
>>         timekeeping_suspended = 0;
>
>
> So using the clocksource-switch test in kselftest, I can pretty easily
> hit inconsistencies with this.
>
> The reason is since we use the coarse_nsec as the nanosecond portion
> of the coarse clockids, I don't think we ever want to set it to zero,
> as whenever we do so, we lose the previous contents and cause the
> coarse time to jump back.

Bah. Obviously. What was I thinking?

> It seems more likely that we'd want to do something similar to
> tk_update_coarse_nsecs() filling it in with the shifted down
> tk->tkr_mono.xtime_nsec.

Indeed. The earlier approach of handing the offset to
timekeeping_update_from_shadow() was exactly doing that. I dropped that
because of the uglyness vs. the TAI update case in adjtimex().

>> +static inline void tk_update_coarse_nsecs(struct timekeeper *tk, u64 offset)
>> +{
>> +       offset *= tk->tkr_mono.mult;
>> +       tk->coarse_nsec = (tk->tkr_mono.xtime_nsec + offset) >> tk->tkr_mono.shift;
>> +}
>
> Thinking more on this, I get that you're providing the offset to save
> the "at the point" time into the coarse value, but I think this ends
> up complicating things.
>
> Instead it seems like we should just do:
>   tk->coarse_nsec = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;

You end up with the same problem again because xtime_nsec can move
backwards when the multiplier is updated, no?

Thanks,

        tglx