Re: [Qemu-devel] Funny -m arguments can crash

[Gleb Natapov <gleb@redhat.com>]

On Mon, Aug 13, 2012 at 04:56:53PM +0300, Avi Kivity wrote:
> On 08/13/2012 04:41 PM, Markus Armbruster wrote:
> > Avi Kivity <avi@redhat.com> writes:
> > 
> >> On 08/08/2012 12:04 PM, Markus Armbruster wrote:
> >>>>
> >>>> Yes please, maybe with a notice to the user.
> >>> 
> >>> Next problem: minimum RAM size.
> >>> 
> >>> For instance, -M pc -m X, where X < 32KiB dies "qemu: fatal: Trying to
> >>> execute code outside RAM or ROM at [...] Aborted (core dumped)" with
> >>> TCG, and "KVM internal error. Suberror: 1" with KVM.
> >>> 
> >>> Should a minimum RAM size be enforced?  Board-specific?
> >>> 
> >>
> >> It's really a BIOS bug causing a limitation of both kvm and tcg to be
> >> hit.  The BIOS should recognize it doesn't have sufficient memory and
> >> hang gracefully (if you can picture that).  It just assumes some low
> >> memory is available and tries to execute it with the results you got.
> > 
> > SeaBIOS indeed assumes it got at least 1MiB of RAM.  It doesn't bother
> > to check CMOS for a smaller RAM size.  However, that bug / feature is
> > currently masked by a QEMU bug: we screw up CMOS contents when there's
> > less than 1 MiB of RAM.  pc_cmos_init():
> > 
> >     int val, nb, i;
> > [...]
> >     /* memory size */
> >     val = 640; /* base memory in K */
> >     rtc_set_memory(s, 0x15, val);
> >     rtc_set_memory(s, 0x16, val >> 8);
> > 
> >     val = (ram_size / 1024) - 1024;
> >     if (val > 65535)
> >         val = 65535;
> >     rtc_set_memory(s, 0x17, val);
> >     rtc_set_memory(s, 0x18, val >> 8);
> > 
> > If ram_size < 1MiB, val goes negative.  Oops.
> > 
> > For instance, with -m 500k, we happily promise 640KiB base memory (CMOS
> > addr 0x15..16), almost 64MiB extended memory (0x17..18 and 0x30..31),
> > yet no memory above 16MiB (0x34..35).
> > 
> > An easy way to fix this is to require 1MiB of RAM :)
> > 
> > But if you like, I'll put sane values in CMOS instead.  That'll expose
> > the SeaBIOS bug.
> 
> IMO we need to fix CMOS reporting.
> 
> (technically we shouldn't touch CMOS NVRAM at all; seabios should
> discover memory size via fwcfg and program it itself.  But it's
> pointless to change it now)
> 
Chipset we emulate does not support all those crazy memory values you
can give to -m.

--
			Gleb.