cvs commit: www/en/releases/6.1R todo.sgml

[Scott Long]

John Baldwin wrote:
> On Thursday 26 January 2006 07:27, Robert Watson wrote:
> 
>>On Thu, 26 Jan 2006, Ceri Davies wrote:
>>
>>>On Thu, Jan 26, 2006 at 09:57:12AM +0000, Murray Stokely wrote:
>>>
>>>>murray      2006-01-26 09:57:12 UTC
>>>>
>>>>  FreeBSD doc repository
>>>>
>>>>  Modified files:
>>>>    en/releases/6.1R     todo.sgml
>>>>  Log:
>>>>  Add kbdmux and sysinstall smp kernel install items from the ideas page
>>>>  to the 6.1 Desired Features list.
>>>
>>>I think it's a little late to mess with sysinstall to that extent for
>>>6.1. Sounds like the kind of thing that could sit in -CURRENT for months,
>>>but hardly anyone would actually be using it.  It seems that the main
>>>problem with sysinstall is that hardly any of our developers use it.
>>>
>>>On to the question: how often does an SMP kernel fail to boot where a UP
>>>one might work?  I remember that this used to be a problem, but if it's
>>>still "too often", can we have just the bits that probe for an mptable
>>>(or however we determine that there is more that one processor) in the UP
>>>kernel without suffering that instability?
>>>
>>>What I'm basically asking is how much of the SMP code is really required
>>>just to detect MP hardware?
>>
>>SMP kernels now pretty much universally run on UP systems, thanks to work
>>John did a couple of years ago.  The problem has historically been a
>>performance once: the overhead of all the atomic instructions to run an SMP
>>kernel on a UP system is significant.  We're working gradually to improve
>>that, but it's still quite noticeable.  There has been talk of run-time
>>compiling/relinking to use different versions of mutexes (and all that),
>>but no progress as far as I know.  I can't speak to how much information
>>the loader has/needs to decide if it should auto-load an SMP kernel.  A
>>simpler version of the world says that you have an SMP kernel in
>>sysinstall, and based on it probing CPUs, it sets the default kernel in the
>>install to GENERIC or SMP.
> 
> 
> Yes, I would very much prefer that the install just use an SMP kernel.  Note 
> that on all the non-i386 architectures we just have SMP on in GENERIC if it 
> is supported.
> 

SMP kernels still do not universally work on all i386 machines.  I know 
that Alpha and Sparc hardware was designed from the ground up to support
SMP, instead of being a bolted on hack like with x86, but that doesn't
change the facts of the situation.  Despite your work, I don't think it
will ever be safe to make SMP be the default on x86.

Scott