bimodal PAE and compatibility

bimodal PAE and compatibility

[John Levon]

We currently ship a PAE 32-bit domU that we can trivially make bimodal,
except that if we set it to "bimodal", then older Xens will default to
thinking the domU is not PAE:

353     dsi->pae_kernel = PAEKERN_no;
354     if ( dsi->__elfnote_section )
355     {
356         p = xen_elfnote_string(dsi, XEN_ELFNOTE_PAE_MODE);
357         if ( p != NULL && strncmp(p, "yes", 3) == 0 )
358             dsi->pae_kernel = PAEKERN_extended_cr3;

Could we have a new note instead?

thanks
john

Re: bimodal PAE and compatibility

[Gerd Hoffmann]

John Levon wrote:
> We currently ship a PAE 32-bit domU that we can trivially make bimodal,
> except that if we set it to "bimodal", then older Xens will default to
> thinking the domU is not PAE:
> 
> 353     dsi->pae_kernel = PAEKERN_no;
> 354     if ( dsi->__elfnote_section )
> 355     {
> 356         p = xen_elfnote_string(dsi, XEN_ELFNOTE_PAE_MODE);
> 357         if ( p != NULL && strncmp(p, "yes", 3) == 0 )
> 358             dsi->pae_kernel = PAEKERN_extended_cr3;
> 
> Could we have a new note instead?

I don't think it improves the situation, for older Xens you need two
kernels anyway, even if the only difference is the content of the
pae_mode elf note ...

cheers,
  Gerd

-- 
Gerd Hoffmann <kraxel@suse.de>

Re: bimodal PAE and compatibility

[John Levon]

On Mon, Jan 08, 2007 at 04:05:36PM +0100, Gerd Hoffmann wrote:

> > We currently ship a PAE 32-bit domU that we can trivially make bimodal,
> > except that if we set it to "bimodal", then older Xens will default to
> > thinking the domU is not PAE:
> > 
> > 353     dsi->pae_kernel = PAEKERN_no;
> > 354     if ( dsi->__elfnote_section )
> > 355     {
> > 356         p = xen_elfnote_string(dsi, XEN_ELFNOTE_PAE_MODE);
> > 357         if ( p != NULL && strncmp(p, "yes", 3) == 0 )
> > 358             dsi->pae_kernel = PAEKERN_extended_cr3;
> > 
> > Could we have a new note instead?
> 
> I don't think it improves the situation, for older Xens you need two
> kernels anyway, even if the only difference is the content of the
> pae_mode elf note ...

What do you mean? Are you saying that Xen complains about an ELF note it
doesn't understand? That would be unfortunate.

john

Re: bimodal PAE and compatibility

[Gerd Hoffmann]

John Levon wrote:
> On Mon, Jan 08, 2007 at 04:05:36PM +0100, Gerd Hoffmann wrote:
> 
>>> We currently ship a PAE 32-bit domU that we can trivially make bimodal,
>>> except that if we set it to "bimodal", then older Xens will default to
>>> thinking the domU is not PAE:
>>>
>>> 353     dsi->pae_kernel = PAEKERN_no;
>>> 354     if ( dsi->__elfnote_section )
>>> 355     {
>>> 356         p = xen_elfnote_string(dsi, XEN_ELFNOTE_PAE_MODE);
>>> 357         if ( p != NULL && strncmp(p, "yes", 3) == 0 )
>>> 358             dsi->pae_kernel = PAEKERN_extended_cr3;
>>>
>>> Could we have a new note instead?
>> I don't think it improves the situation, for older Xens you need two
>> kernels anyway, even if the only difference is the content of the
>> pae_mode elf note ...
> 
> What do you mean? Are you saying that Xen complains about an ELF note it
> doesn't understand?

No.

Moving the bimodal flag to another elf node doesn't fix the problem that
older Xen versions don't support bimodal kernels, thus you still need a
pae and a non-pae kernel.  As older Xens consider bimodal kernels as
non-pae one bimodal and one pae kernel works too.  But you can't get
that down to only one kernel unless you drop support for xen versions
below 3.0.4 ...

cheers,
  Gerd

-- 
Gerd Hoffmann <kraxel@suse.de>

Re: bimodal PAE and compatibility

[John Levon]

On Mon, Jan 08, 2007 at 04:24:04PM +0100, Gerd Hoffmann wrote:

> Moving the bimodal flag to another elf node doesn't fix the problem that
> older Xen versions don't support bimodal kernels

I never said it did. What it /does/ do is allow you to specify a correct
fallback rather than assuming that pae=no is the right one. It isn't for
us: currently we do not intend to ship a non-PAE kernel as we have no
infrastructure for such an (alien to Solaris) restriction.

> non-pae one bimodal and one pae kernel works too.  But you can't get
> that down to only one kernel unless you drop support for xen versions
> below 3.0.4 ...

Or drop any support for non-PAE before 3.0.4 which is exactly what we
were and are doing anyway.

regards
john

Re: bimodal PAE and compatibility

[Keir Fraser]

On 8/1/07 14:49, "John Levon" <levon@movementarian.org> wrote:

> 353     dsi->pae_kernel = PAEKERN_no;
> 354     if ( dsi->__elfnote_section )
> 355     {
> 356         p = xen_elfnote_string(dsi, XEN_ELFNOTE_PAE_MODE);
> 357         if ( p != NULL && strncmp(p, "yes", 3) == 0 )
> 358             dsi->pae_kernel = PAEKERN_extended_cr3;
> 
> Could we have a new note instead?

I'd take a patch to strstr for 'bimodal' instead of strncmping for it. Then
you can put something like 'yes,bimodal' in your elf note.

 -- Keir

Re: bimodal PAE and compatibility

[Gerd Hoffmann]

  Hi,

> Or drop any support for non-PAE before 3.0.4 which is exactly what we
> were and are doing anyway.

Ok, that is another way ...

cheers,
  Gerd

-- 
Gerd Hoffmann <kraxel@suse.de>

Re: bimodal PAE and compatibility

[John Levon]

On Mon, Jan 08, 2007 at 04:24:35PM +0000, Keir Fraser wrote:

> I'd take a patch to strstr for 'bimodal' instead of strncmping for it. Then
> you can put something like 'yes,bimodal' in your elf note.

Good idea, thanks. I'll have to get Xen to compile again first though :)

john