linux-kernel - Re: [PATCH v2 3/7] timecounter: allow for non-power of two overflow

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] [day] [month] [year] [list]

Message-ID: <874k81l1db.ffs@tglx>
Date:   Wed, 24 Nov 2021 21:55:44 +0100
From:   Thomas Gleixner <tglx@...utronix.de>
To:     Martin Kaistra <martin.kaistra@...utronix.de>,
        Florian Fainelli <f.fainelli@...il.com>,
        Andrew Lunn <andrew@...n.ch>,
        Vivien Didelot <vivien.didelot@...il.com>
Cc:     martin.kaistra@...utronix.de,
        Richard Cochran <richardcochran@...il.com>,
        Vladimir Oltean <olteanv@...il.com>,
        "David S. Miller" <davem@...emloft.net>,
        Jakub Kicinski <kuba@...nel.org>,
        John Stultz <john.stultz@...aro.org>,
        Stephen Boyd <sboyd@...nel.org>,
        Russell King <linux@...linux.org.uk>,
        Marc Kleine-Budde <mkl@...gutronix.de>,
        linux-kernel@...r.kernel.org, netdev@...r.kernel.org
Subject: Re: [PATCH v2 3/7] timecounter: allow for non-power of two overflow

Martin,

On Tue, Nov 09 2021 at 10:50, Martin Kaistra wrote:
>   *			see CYCLECOUNTER_MASK() helper macro
>   * @mult:		cycle to nanosecond multiplier
>   * @shift:		cycle to nanosecond divisor (power of two)
> + * @overflow_point:	non-power of two overflow point (optional),
> + *			smaller than mask
>   */
>  struct cyclecounter {
>  	u64 (*read)(const struct cyclecounter *cc);
>  	u64 mask;
>  	u32 mult;
>  	u32 shift;
> +	u64 overflow_point;
>  };
>  
>  /**
> diff --git a/kernel/time/timecounter.c b/kernel/time/timecounter.c
> index e6285288d765..afd2910a9724 100644
> --- a/kernel/time/timecounter.c
> +++ b/kernel/time/timecounter.c
> @@ -39,6 +39,9 @@ static u64 timecounter_read_delta(struct timecounter *tc)
>  	/* calculate the delta since the last timecounter_read_delta(): */
>  	cycle_delta = (cycle_now - tc->cycle_last) & tc->cc->mask;
>  
> +	if (tc->cc->overflow_point && (cycle_now - tc->cycle_last) > tc->cc->mask)
> +		cycle_delta -= tc->cc->mask - tc->cc->overflow_point;

TBH, this took me more than one twisted braincell to grok.

With support for clocks which do not wrap at power of 2 boundaries we
already lose the unconditional fast path no matter what. So what's the
point of having two conditions and doing this convoluted math here?

In timecounter_init():

   	tc->ovfl = cc->ovfl ? cc->ovfl : cc->mask + 1;

which makes it a common path in timecounter_read_delta():

  	cycle_delta = cycle_now - tc->cycle_last;
        if ((s64)cycle_delta) < 0)
        	cycle_delta += tc->ovfl;

which produces way better binary code.

The conditional does not really matter for the timecounter use cases as
that calculation is noise compared to the actual cc->read() access.

Aside of that the same problem exists in timecounter_cyc2time()...

After that we probably should do a treewide sweep to get rid of cc->mask
to avoid confusion and subtle to understand errors when some code uses
cc->mask instead of cc->ovfl.

Thanks,

        tglx