[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] Limitation in HVM physmap



On Fri, 18 Oct 2013, Ian Campbell wrote:
> On Fri, 2013-10-18 at 15:56 +0100, Wei Liu wrote:
> > On Fri, Oct 18, 2013 at 04:41:49PM +0200, Tim Deegan wrote:
> > > At 15:36 +0100 on 18 Oct (1382107000), Wei Liu wrote:
> > > > On Fri, Oct 18, 2013 at 04:28:24PM +0200, Tim Deegan wrote:
> > > > > Hi,
> > > > > 
> > > > > At 15:20 +0100 on 18 Oct (1382106012), Wei Liu wrote:
> > > > > > I currently run into the limitation of HVM's physmap: one MFN can 
> > > > > > only
> > > > > > be mapped into one guest physical frame. Why is it designed like 
> > > > > > that?
> > > > > 
> > > > > 1. For simplicity.  That code is hard wnough to work with already. :)
> > > > > 
> > > > > 2. It helps avoid worrying about inconsistent cache settings if the
> > > > >    HAP tables only have one entry for each MFN. 
> > > > > 
> > > > > 3. Xen maintains a single mapping from MFN back to PFN, and any code
> > > > >    that uses it would have to be able to deal with getting multiple
> > > > >    answers.  That's already been partly changed by the mem_sharing
> > > > >    code (which obviously _can_ have multiple P2M entries pointing to
> > > > >    the same MFN but is a bit complex as a result).
> > > > > 
> > > > 
> > > > I see.
> > > > 
> > > > > > The scenario is that: when QEMU boots with OVMF (UEFI firmware), 
> > > > > > OVMF
> > > > > > will first map the framebuffer to 0x80000000, resulting the 
> > > > > > framebuffer
> > > > > > MFNs added to corresponding slots in physmap. A few moments later 
> > > > > > when
> > > > > > Linux kernel loads, it tries to map framebuffer MFNs to 0xf00000000,
> > > > > > which fails because those MFNs have already been mapped in other
> > > > > > locations. Is there a way to fix this?
> > > > > 
> > > > > Qemu could remember where it put the framebuffer last time and unmap
> > > > > it.  AIUI that's how real hardware would behave if you changed the
> > > > > framebuffer base address -- the framebuffer wouldn't still be mapped 
> > > > > at
> > > > > the old location as well.
> > > > > 
> > > > 
> > > > In fact, this is not the case in OVMF. EFIFB driver in Linux will always
> > > > use the EFIFB region (0x80000000) provided by OVMF. 
> > > 
> > > So what's mapping it at 0xf00000000?  Is linux running two drivers
> > > against the same graphics card at the same time?  That sounds like
> > > trouble!
> > > 
> > 
> > During OVMF initialization:
> > 
> > (d28) PciBus: Resource Map for Root Bridge PciRoot(0x0)                     
> >     
> > (d28) Type =   Io16; Base = 0xC000;     Length = 0x1000;        Alignment = 
> > 0xFFF
> > (d28)  Base = 0xC000;   Length = 0x100; Alignment = 0xFF;       Owner = PCI 
> >  [00|04|
> > (d28)  Base = 0xC100;   Length = 0x100; Alignment = 0xFF;       Owner = PCI 
> >  [00|03|
> > (d28)  Base = 0xC200;   Length = 0x10;  Alignment = 0xF;        Owner = PCI 
> >  [00|01|
> > (d28) Type =  Mem32; Base = 0x80000000; Length = 0x3100000;     Alignment = 
> > 0x1FFFFF
> > (d28)  Base = 0x80000000;       Length = 0x2000000;     Alignment = 
> > 0x1FFFFFF;  Owne
> > (d28) |02|00:10] 
> > 
> > Later when Linux loads EFIFB driver:
> > [    2.628264] efifb: framebuffer at 0x80000000, mapped to
> > 0xffffc90000100000, using 1876k, total 1875k
> > [    2.646827] efifb: mode is 800x600x32, linelength=3200, pages=1
> > [    2.658833] efifb: scrolling: redraw
> > [    2.666342] efifb: Truecolor: size=8:8:8:8, shift=24:16:8:0
> > 
> > 0xf0000000 is mapped by:
> 
> What hardware is backing the framebuffer at 0x80000000? It doesn't
> appear to be this cirrus device.

Is it possible that the UEFI VGA framebuffer is exposed via two
different regions?
After all that is what happens today on any VGA card: one is the PCI
MMIO region, the other is the legacy VGA framebuffer address.
Maybe 0x80000000 is the VGA framebuffer configured by the UEFI firmware
and needs to stay accessible even after the driver sets the PCI BAR for
the card.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.