Xen devel - Aug 2008 - Dom heap virtual address question

Silly x86-related memory question from someone who has only
used global virtual address machines for too many years:

Is it possible to take the result of an alloc_domheap_pages()
(struct page_info *) and convert this to a virtual address
that can be used at a later time in the hypervisor (not
in a domain)?  E.g.

va = mfn_to_virt(page_to_mfn(alloc_domheap_pages()))
  :
(many seconds later, still in hypervisor)
  :
a = *(unsigned int *)va;

Or is it necessary to do a va = map_domain_page(mfn)
then unmap_domain_page(va) around each usage?  (And
if so do I need the _global version of these?)

I''m willing to restrict this to 64-bit hypervisor only
if that matters...  but ideally I''d like it to be
portable (i.e. to ia64/ppc).

Also, if it matters, I''d probably use anonymous domain
(NULL) rather than current->domain.

Thanks,
Dan

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

>From: Daniel Magenheimer
>Sent: 2008年8月27日 7:49
>
>Silly x86-related memory question from someone who has only
>used global virtual address machines for too many years:
>
>Is it possible to take the result of an alloc_domheap_pages()
>(struct page_info *) and convert this to a virtual address
>that can be used at a later time in the hypervisor (not
>in a domain)?  E.g.
>
>va = mfn_to_virt(page_to_mfn(alloc_domheap_pages()))
>  :
>(many seconds later, still in hypervisor)
>  :
>a = *(unsigned int *)va;
This only applies to 64bit xen which directly maps all memory 
in a large enough virtual address range. For 32bit xen, mfn_to_virt
only applies to xenheap pages.
>
>Or is it necessary to do a va = map_domain_page(mfn)
>then unmap_domain_page(va) around each usage?  (And
>if so do I need the _global version of these?)
map/unmap_domain_page is cleaner wrap, suitable for both
32bit and 64bit xen. For 64bit xen, it''s actually a macro to
mfn_to_virt and thus map/unmap is light. you can use _global
version if you''d like to access that page from any domain 
context and don''t want to local map/unmap frequently.

Thanks,
Kevin

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Hi Kevin --

Thanks much for the reply.

Further clarification on the _global versions:
If domainA makes a hypercall that results in the

va = mfn_to_virt(page_to_mfn(alloc_domheap_pages()))

(or equivalent wrapped in map/unmap_domain_page)

and then domainB makes a hypercall but IN THE
HYPERVISOR I want to make use of the previously
saved va (or mapped mfn), is that still in the context
of domainA?  In other words, I must use _global?

Thanks,
Dan
> -----Original Message-----
> From: Tian, Kevin [mailto:kevin.tian@intel.com]
> Sent: Tuesday, August 26, 2008 6:35 PM
> To: Daniel Magenheimer; Xen-Devel (E-mail)
> Subject: RE: [Xen-devel] Dom heap virtual address question
>
>
> >From: Daniel Magenheimer
> >Sent: 2008年8月27日 7:49
> >
> >Silly x86-related memory question from someone who has only
> >used global virtual address machines for too many years:
> >
> >Is it possible to take the result of an alloc_domheap_pages()
> >(struct page_info *) and convert this to a virtual address
> >that can be used at a later time in the hypervisor (not
> >in a domain)?  E.g.
> >
> >va = mfn_to_virt(page_to_mfn(alloc_domheap_pages()))
> >  :
> >(many seconds later, still in hypervisor)
> >  :
> >a = *(unsigned int *)va;
>
> This only applies to 64bit xen which directly maps all memory
> in a large enough virtual address range. For 32bit xen, mfn_to_virt
> only applies to xenheap pages.
>
> >
> >Or is it necessary to do a va = map_domain_page(mfn)
> >then unmap_domain_page(va) around each usage?  (And
> >if so do I need the _global version of these?)
>
> map/unmap_domain_page is cleaner wrap, suitable for both
> 32bit and 64bit xen. For 64bit xen, it''s actually a macro to
> mfn_to_virt and thus map/unmap is light. you can use _global
> version if you''d like to access that page from any domain
> context and don''t want to local map/unmap frequently.
>
> Thanks,
> Kevin
>
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xensource.com
> http://lists.xensource.com/xen-devel
>
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

>From: Daniel Magenheimer [mailto:dan.magenheimer@oracle.com] 
>Sent: 2008年8月27日 9:20
>
>Hi Kevin --
>
>Thanks much for the reply.
>
>Further clarification on the _global versions:
>If domainA makes a hypercall that results in the
>
>va = mfn_to_virt(page_to_mfn(alloc_domheap_pages()))
>
>(or equivalent wrapped in map/unmap_domain_page)
>
>and then domainB makes a hypercall but IN THE
>HYPERVISOR I want to make use of the previously
>saved va (or mapped mfn), is that still in the context
>of domainA?  In other words, I must use _global?
>
Yes. The ptes used for non global version is domain
specific, which is changed along with context switch.
However ptes for _global version are persistent across
domains. This is for 32bit xen.

For 64bit xen, both local and global version are simply
macros to mfn_to_virt which is a directly mapping
style setuped at boot time. In such case, you can always 
access previously mapped va from any domain context, 
regardless of which interface is used.

But for portability, it''s better to have 32bit implication
in mind to avoid mistake. :-)

Thanks,
Kevin

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

> But for portability, it''s better to have 32bit implication
> in mind to avoid mistake. :-)
Yes, exactly what I was thinking!
> For 64bit xen, both local and global version are simply
> macros to mfn_to_virt which is a directly mapping
OK, I found this in xen/include/xen/domain_page.h
but was confused because CONFIG_DOMAIN_PAGE must
be set to zero and include/asm-x86/config.h sets
it to one.  But apparently the 64-bit build scripts
override the config.h setting and CONFIG_DOMAIN_PAGE=1
is used.

Thanks again!
Dan
> -----Original Message-----
> From: Tian, Kevin [mailto:kevin.tian@intel.com]
> Sent: Tuesday, August 26, 2008 7:29 PM
> To: Daniel Magenheimer; Xen-Devel (E-mail)
> Subject: RE: [Xen-devel] Dom heap virtual address question
>
>
> >From: Daniel Magenheimer [mailto:dan.magenheimer@oracle.com]
> >Sent: 2008年8月27日 9:20
> >
> >Hi Kevin --
> >
> >Thanks much for the reply.
> >
> >Further clarification on the _global versions:
> >If domainA makes a hypercall that results in the
> >
> >va = mfn_to_virt(page_to_mfn(alloc_domheap_pages()))
> >
> >(or equivalent wrapped in map/unmap_domain_page)
> >
> >and then domainB makes a hypercall but IN THE
> >HYPERVISOR I want to make use of the previously
> >saved va (or mapped mfn), is that still in the context
> >of domainA?  In other words, I must use _global?
> >
>
> Yes. The ptes used for non global version is domain
> specific, which is changed along with context switch.
> However ptes for _global version are persistent across
> domains. This is for 32bit xen.
>
> For 64bit xen, both local and global version are simply
> macros to mfn_to_virt which is a directly mapping
> style setuped at boot time. In such case, you can always
> access previously mapped va from any domain context,
> regardless of which interface is used.
>
> But for portability, it''s better to have 32bit implication
> in mind to avoid mistake. :-)
>
> Thanks,
> Kevin
>
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel