Re: [RFC] xen/x86: allow overlaps with non-RAM regions

[Jason Andryuk <jason.andryuk@xxxxxxx>]

On 2025-04-24 06:48, Roger Pau Monné wrote:
> On Thu, Apr 24, 2025 at 12:15:17PM +0200, Marek Marczykowski-Górecki wrote:
> > On Thu, Apr 24, 2025 at 09:59:00AM +0200, Roger Pau Monné wrote:
> > > On Wed, Apr 23, 2025 at 04:51:16PM -0700, Lira, Victor M wrote:
> > > > [    4.570354] Intel(R) 2.5G Ethernet Linux Driver
> > > >
> > > > [    4.579535] Copyright(c) 2018 Intel Corporation.
> > > >
> > > > [    4.588898] sky2: driver version 1.30
> > > >
> > > > (XEN) [   21.644361] d0v3 unable to fixup memory read from 0xfe91000c size 4: -1
> > > This fault is unexpected, according to the identity mappings done at
> > > domain build time:
> > >
> > > d0: identity mappings for IOMMU:
> > >   [00000000a0, 00000000ff] RW
> > >   [0000009bff, 0000009fff] RW
> > >   [00000cabc9, 00000cc14c] RW
> > >   [00000cc389, 00000cc389] RW
> > >   [00000cc70a, 00000cd1fe] RW
> > >   [00000ce000, 00000cffff] RW
> > >   [00000fd000, 00000fd2ff] RW
> > >
> > > The page at 0xfe910 should be covered by the last range above.  I
> > > think we have a bug somewhere that unexpectedly unmaps that address.
> > You sure? 0xfe910 is outside of [00000fd000, 00000fd2ff].
> Oh, did and off-by-one when copying, it should have been:
>
> d0: identity mappings for IOMMU:
>   [00000000a0, 00000000ff] RW
>   [0000009bff, 0000009fff] RW
>   [00000cabc9, 00000cc14c] RW
>   [00000cc389, 00000cc389] RW
>   [00000cc70a, 00000cd1fe] RW   
>   [00000ce000, 00000cffff] RW
>   [00000fd000, 00000fd2ff] RW
>   [00000fd304, 00000febff] RW
>
> Where 0xfe910 is covered by the last range.
This looks like it's okay.  You think the somehow the device ranges are 
unmapped?
I thought the immediate issue is that the e820 marks these addresses 
reserved.  That leads to pci_check_bars()->is_memory_hole() returning 
false.  Stefano's original patch just returned true for non-RAM memory 
regions and the device worked.  That would leave bar->mem rangeset 
empty, IIUC.  Not sure how that would remove mappings though.
That also changes the setting of bar->enabled, but I don't immediately 
see how that affects mappings.
Regards,
Jason