linux-kernel - Re: [PATCH] sched: fix/optimise calculation of weight-inverse

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]

Date:	Wed, 11 May 2011 18:20:30 +0200
From:	Ingo Molnar <mingo@...e.hu>
To:	Stephan Bärwolf 
	<stephan.baerwolf@...ilmenau.de>
Cc:	Linus Torvalds <torvalds@...ux-foundation.org>,
	Nikhil Rao <ncrao@...gle.com>,
	Peter Zijlstra <peterz@...radead.org>,
	Mike Galbraith <efault@....de>,
	"Nikunj A. Dadhania" <nikunj@...ux.vnet.ibm.com>,
	Srivatsa Vaddagiri <vatsa@...ux.vnet.ibm.com>,
	linux-kernel@...r.kernel.org
Subject: Re: [PATCH] sched: fix/optimise calculation of weight-inverse


* Stephan Bärwolf <stephan.baerwolf@...ilmenau.de> wrote:

> If the inverse loadweight should be zero, function "calc_delta_mine"
> calculates the inverse of "lw->weight" (in 32bit integer ops).
> 
> This calculation is actually a little bit impure (because it is
> inverting something around "lw-weight"+1), especially when
> "lw->weight" becomes smaller. (This could explain some aritmetical
> issues for small shares...)
> 
> The correct inverse would be 1/lw->weight multiplied by
> "WMULT_CONST" for fixcomma-scaling it into integers.
> (So WMULT_CONST/lw->weight ...)
> 
> For safety it is also important to check if division by zero
> could happen...
> 
> The old, impure algorithm took two divisions for inverting lw->weight,
> the new, more exact one only takes one and an additional unlikely-if.
> 
> Signed-off-by: Stephan Baerwolf <stephan.baerwolf@...ilmenau.de>
> ---
>  kernel/sched.c |   12 +++++++++---
>  1 files changed, 9 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/sched.c b/kernel/sched.c
> index 312f8b9..bb55996 100644
> --- a/kernel/sched.c
> +++ b/kernel/sched.c
> @@ -1307,15 +1307,21 @@ calc_delta_mine(unsigned long delta_exec,
> unsigned long weight,
>  {
>      u64 tmp;
>  
> +    tmp = (u64)delta_exec * weight;
> +
> +    // actually we would have to trap - division by zero - but we stay
> and pretend the limit of the operation...
> +    if (unlikely(lw->weight == 0)) {
> +        if (unlikely(tmp == ((u64)0))) return (unsigned long)0;
> +        else return (unsigned long)LONG_MAX;

Can lw->weight ever be zero here? I dont think so - and if it is then getting a 
kernel crash there is preferred to hiding it.

Once we do that your patch becomes a lot simpler.

> +    }
> +
>      if (!lw->inv_weight) {
>          if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
>              lw->inv_weight = 1;
>          else
> -            lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
> -                / (lw->weight+1);
> +            lw->inv_weight = WMULT_CONST / lw->weight;

hm, i definitely think there was a rounding reason for that - but apparently 
i'm an idiot who does not add comments to non-obvious code! :-)

Peter, do you remember this?

>      }
>  
> -    tmp = (u64)delta_exec * weight;

I agree that moving this multiplication early in the sequence is better for 
micro-performance regardless of the lw->weight optimization you do: it can be 
executed in parallel.

Thanks,

	Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/