Xen devel - Dec 2013 - dom0 pvh: linux: issues without dom0_mem

Konrad,

Looks like there are bunch of things in the xen_set_identity_and_release
code for pvh. Things crap out for dom0 when dom0_mem is not specified.

For one, the code in  xen_pvh_adjust_stats():

                unsigned long end = min(max_pfn_mapped, end_pfn);

is wrong. The result is a negative number, and causes released to be
set improperly:

[    0.000000] Released 18446744073707253146 pages of unused memory

This then later causes xen_do_chunk to barf:

[    0.000000] ------------[ cut here ]------------
[    0.000000] WARNING: CPU: 0 PID: 0 at
/home/sb/hybrid/linux-v3.12/arch/x86/xen/setup.c:134 xen_do_chunk+0x1a0/0x247()
[    0.000000] Failed to populate pfn 271f85 err=0


Second, I don''t remember or understand why we removed the
xen_release_chunk
from pvh path, and now just collect the stats. If anything is mapped, it 
would need to be removed. Oh never mind, i remember, because when the iomap
is done by xen, it has already removed those pages. Hmm... so let me
figure how to fix xen_pvh_adjust_stats(). I''ll send patch. JFYI you and
others.

thanks
Mukesh

On Tue, 3 Dec 2013 18:20:40 -0800
Mukesh Rathor <mukesh.rathor@oracle.com> wrote:
> Konrad,
> 
> Looks like there are bunch of things in the
> xen_set_identity_and_release code for pvh. Things crap out for dom0
> when dom0_mem is not specified.
> 
> For one, the code in  xen_pvh_adjust_stats():
> 
>                 unsigned long end = min(max_pfn_mapped, end_pfn);
> 
> is wrong. The result is a negative number, and causes released to be
> set improperly:
> 
> [    0.000000] Released 18446744073707253146 pages of unused memory
> 
> This then later causes xen_do_chunk to barf:
> 
> [    0.000000] ------------[ cut here ]------------
> [    0.000000] WARNING: CPU: 0 PID: 0
> at /home/sb/hybrid/linux-v3.12/arch/x86/xen/setup.c:134
> xen_do_chunk+0x1a0/0x247() [    0.000000] Failed to populate pfn
> 271f85 err=0
> 
> 
> Second, I don''t remember or understand why we removed the
> xen_release_chunk from pvh path, and now just collect the stats. If
> anything is mapped, it would need to be removed. Oh never mind, i
> remember, because when the iomap is done by xen, it has already
> removed those pages. Hmm... so let me figure how to fix
> xen_pvh_adjust_stats(). I''ll send patch. JFYI you and others.
> 
Tim/Jan,

With regard to this, when a PVH maps all iomem upfront, the old mfn''s
are
discarded. I would have thought, and I think I had that code orginally
somewhere, that set_mmio_p2m_entry would free_domheap_page() at:

    else if ( p2m_is_ram(ot) )
    {
        ASSERT(mfn_valid(omfn));
        set_gpfn_from_mfn(mfn_x(omfn), INVALID_M2P_ENTRY);
          NO free_domheap_page(omfn) <---------
    }

When dom0_mem is not specified, and lotsa pages are freed, that would
be a big memleak. If it can''t be done here, then I could add something
to pvh_add_mem_mapping() in domain_build.c (patch 2 of v4 dom0 series):

static __init void pvh_add_mem_mapping(struct domain *d, unsigned long gfn,
                                       unsigned long mfn, unsigned long nr_mfns)
{
    unsigned long i;
    for ( i = 0; i < nr_mfns; i++ )

        if (old mfn && it''s type is RAM)
            free_domheap_page(oldmfn);

        if ( !set_mmio_p2m_entry(d, gfn + i, _mfn(mfn + i)) )
            panic("Failed setting p2m. gfn:%lx mfn:%lx i:%ld\n", gfn,
mfn, i);
}

Thoughts?

thanks
mukesh