Re: [PATCHES] 686-bigmem/Xen enabled netboot images
On Sat, 2008-07-12 at 19:25 +0200, Frans Pop wrote:
> On Thursday 10 July 2008, Ian Campbell wrote:
> > First patch is to kernel wedge and adds the Xen block and net devices
> > (optional since they won't appear in the 486 images) as well as making
> > generic_serial optional in order to allow 686-bigmem kernel udebs to be
> > built. [kernel-wedge.patch]
>
> ! +++ packages/kernel/kernel-wedge/modules/nic-modules (working copy)
> ! +xen-netfront ?
> ! +++ packages/kernel/kernel-wedge/modules/scsi-modules (working copy)
> ! +xen-blkfront ?
>
> As these modules are only going to be used with the i386 -bigmem kernel
> and even only exist there, I wonder if we want them in kernel-wedge.
They will eventually be needed for the amd64 kernels too, although
probably not in time for Lenny. There's also a port of Xen to IA64 which
is fairly active and just starting to go into the upstream kernel, again
probably not for Lenny though.
What about retasking the new virtio udebs into virtuali{s,z}ion (or just
virt) and adding them there?
(aside: how does the installer go from the presence of some piece of
hardware, virtual or otherwise, to a dependency on a particular udeb??)
> In this case I think adding them only where they are actually used is
> preferable. That would mean adding them in relevant files in
> linux-kernel-di-i386-2.6/modules/i386/ instead.
>
> > Second patch is to linux-kernel-di-i386-2.6 and simply adds the
> > 686-bigmem flavour kernels. [linux-kernel-di-i386-2.6.patch]
>
> I think a bit longer explanation in the changelog would be good: that the
> new variant is intended to be used for Xen installations.
Sure. Now reads:
[ Ian Campbell ]
* Build -686-bigmem udebs, these are needed for the installer variant which
runs on PAE and 64 bit Xen hypervisors.
> > Third patch is to base-installer and causes the 686-bigmem kernel to be
> > installed into the new system iff the installer is also running a
> > bigmem kernel. This has been filed as #480054 and I'll send an update
> > there too. [base-installer.patch]
>
> First of all I'm afraid that even your latest patch is going to be
> invalidated by a new change in i386 kernel selection: #490542 :-P
I saw it and suspected as much, no worries...
> But more importantly, the last hunk of the patch is not going to do what
> you want. Did you run the testsuite ('cd kernel/tests; ./runtests i386')?
> I think you'll see failures. Unless I'm very much mistaken your code does
> not produce the desired fallback to first the "plain" -686 flavor.
You are right -- I noticed it later and changed it locally (and wrongly
now I think about it...) but forgot to resend... I'll fix up after
490542 goes in and resend then.
> As #490542 is a further simplification I think it's easiest to have your
> change on top of that one (I expect to commit tomorrow). And with those
> changes I don't think a separate testcase for AMD is still needed.
>
> > Final patch is the the installer itself to cause a 686-bigmem netboot
> > image to be built. [installer.patch]
>
> ! +++ installer/build/config/i386/netboot-bigmem.cfg (revision 0)
>
> AFAICT the following is missing to ensure the result lands in a subdir:
> EXTRANAME = netboot/bigmem/
It currently ends up in installer/build/dest/netboot-bigmem/ we want
installer/build/dest/netboot/bigmem? That's fine by me (modulo naming
discussion below).
> Besides that I still wonder if just "bigmem" is the best name for this. I
> would think that users are going to be looking for Xen images.
>
> My proposal would be to use "netboot-xen" as internal name (the fact that
> we need the bigmem kernel for Xen is secondary to the purpose for which
> we build it IMO) and maybe "netboot/xen-bigmem" for the directory
> (EXTRANAME).
I really wanted to avoid having a Xen specific image, but you are right
it will most likely confuse users. I don't suppose there is any support
for aliases (becoming symlinks or something in the output)? I could add
that functionality if desired. If not then I'd say either netboot/xen or
netboot/xen-bigmem would be an appropriate name.
Other distros (and *bsd etc) would tend to call it either xen or xenpae,
I think, rather than bigmem (which is something of a Debianism) -- what
do you think of those names?
>
> > New from last time is a patch to finish-install to add a getty on hvc0
> > if Xen is detected. It's unfortunate that the test for Xen has to be
> > present but this is the only way which I could find that work reliably.
> > [finish-install.patch]
>
> No real problems except the indentation: please use tabs as in the rest of
> the file.
Done.
> BTW, I propose that you (if you're interested) are given commit access to
> the D-I repository on alioth so you can commit this yourself and maintain
> Xen (and maybe virtualization in general) support directly. I see you
> already have an account.
That's fine by me although I would still like to keep posting to the
list first (and you probably want me to anyway). My alioth account is
ijc-guest.
I'll prep a new set of patches and resend once I see your base-installer
changes.
Ian.
--
Ian Campbell
Why are you so hard to ignore?
Reply to: