Re: CVS commit: src/sys/kern

[Nick Hudson <nick.hudson%gmx.co.uk@localhost>]

On 25/08/2018 14:50, Maxime Villard wrote:
> Le 25/08/2018 à 14:37, Nick Hudson a écrit :
> > > > On 25 Aug 2018, at 10:58, Maxime Villard <max%m00nbsd.net@localhost> wrote:
> > > > > Le 25/08/2018 à 11:50, Nick Hudson a écrit :
> > > > > On 25/08/2018 06:56, Maxime Villard wrote:
> > > > > Module Name:    src
> > > > > Committed By:    maxv
> > > > > Date:        Sat Aug 25 05:56:25 UTC 2018
> > > > >
> > > > > Modified Files:
> > > > >    src/sys/kern: subr_pool.c
> > > > >
> > > > > Log Message:
> > > > > Disable POOL_REDZONE until we figure out what's wrong. There must 
> > > > > be a dumb
> > > > > problem, that is not triggerable on amd64.
> > > >
> > > > Reversed align and align_offset arguments to pool_cache_bootstrap 
> > > > appear to be
> > > > the problem.
> > >
> > > indeed
> >
> > I rushed this email out... I think pool red zone needs fixing for non 
> > zero
> > align_offset?
>
> I don't think so, unless you see a specific problem.
>
> pool_redzone_init() only modifies pr_size. align_offset is touched 
> afterwards,
> depending on pr_size, and I don't see why we would need to take care 
> of it.
>
> I hardly see how the inverted parameters can be correct. Already in 
> the code
> we compute (align - ioff), so if you pass align < ioff, we have a 
> problem.

Let me try this again...

There is clearly a problem with POOL_REDZONE and align < align_offset. 
Currently subr_pool.c has

585:         * Silently enforce `0 <= ioff < align'.
586:         */
587:        pp->pr_itemoffset = ioff %= align;

Either this is made loud or POOL_REDZONE is fixed - which do you prefer?

Thanks,
Nick