On 4/13/06, Eric W. Biederman <[email protected]> wrote:
> "Magnus Damm" <[email protected]> writes:
>
> > On 4/13/06, Eric W. Biederman <[email protected]> wrote:
>
> >> Feel free to fix x86_64, to use only page sized allocates.
> >
> > I will. But first - questions:
> >
> > Should KEXEC_CONTROL_CODE_SIZE be left in even if it's always 4096?
>
> So far I don't see a compelling case to remove it. To a certain
> extent I am happily surprised to see that everyone's code
> across several architectures has managed to fit in 4KB.
I'm happy about that fact too. So I will not remove that constant then.
> > Do you like how I added image->arch_private?
>
> At a quick glance I couldn't make sense of the interactions.
> So I totally missed that.
>
> So you actually did 3 things at once. That makes a very hard
> to digest patch.
Ok, I will break out the x86_64 stuff and resend.
> >> Until I see a reasonable argument that none of the architectures
> >> currently supported by the linux kernel would need a multi order
> >> allocation for a kexec port am I interested in removing support.
> >
> > I argue that it is quite pointless to have code to support N-order
> > allocations that no one is using. Especially since the code is more
> > complex and it may be harder for the buddy allocator to fulfill
> > N-order allocations compared to 0-order allocations.
>
> The complexity as your patch shows is currently is 2 for loops.
> Refactoring the entire code base to save 2 for loops when
> using N-order allocations are totally voluntary is over kill.
>
> Most of the complexity in the code actually comes from having
> to use 0-order allocations.
I'm not saying that the kexec code itself should be made simpler, I
just think it is bad to keep unused code left in the source tree. And
regarding the complexity from 0-order allocations - yes - it becomes
more complex to handle single pages but that's what you have to pay
for to avoid fragmentation, right?
My argument against keeping support for N-order allocations is that
someone might feel that it is a good solution to allocate contiguous
pages to workaround something (like x86_64 does) instead of only using
N-order allocations for situations where physically contiguous pages
really are required by hardware.
Non-0-order allocations should be avoided as much as possible IMO.
This is somewhat related to the 4K vs 8K stack discussion, but a much
less frequent allocation problem of course.
> > And on top of the reasons above I'd like to stay away from N-order
> > allocations because Xen doesn't guarantee that (pseudo-)physical pages
> > handled out by the buddy allocator are contiguous.
>
> Yes. Xen doesn't have enough sense to use 4MB pages so kernels can
> execute efficiently. That may be overly harsh. But given the
> efficiency that you can get from using large pages in the kernel
> not guaranteeing large page allocations seems quite foolish.
Many things can be said about Xen. Maybe they use the keep-it-simple
strategy to begin with. I'm quite sure it will be handled (or at least
will be put on a road map) if it significantly improves performance.
> >> As I recall the alpha had an architectural need for a 32KB
> >> allocation or something like that.
> >
> > Oh. So if someone is working on kexec for alpha I guess we need
> > N-order allocations, right?
>
> To be clear. Until some one shows me that on no architecture
> that the linux kernel supports there are no data structures
> that the cpus use directly that exceed 1 page there is the potential
> to need > 0 order allocations.
>
> My investigation into the basic problem says the are occasions
> when order-N allocations are needed.
I'm absolutely sure you are right about that. But keeping unused code
in kernel is another question IMO.
> I am overjoyed that currently there are multiple architectures
> supported by kexec but the porting work has yet to slow
> down as your Xen work shows.
>
> kexec currently does not have the volume of development and the number
> of people who understand it to handle being refactored very often.
> For preparation phase we are likely ok. For later when it gets into
> very tricky arch specific assembly things are much worse. Unless the
> basic skeleton gets it the way please use what is provided has been
> debugged.
Ok, I will avoid modifying the generic framework then.
Thanks for the detailed explanation.
/ magnus
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[Index of Archives]
[Kernel Newbies]
[Netfilter]
[Bugtraq]
[Photo]
[Stuff]
[Gimp]
[Yosemite News]
[MIPS Linux]
[ARM Linux]
[Linux Security]
[Linux RAID]
[Video 4 Linux]
[Linux for the blind]
[Linux Resources]
