Re: [Qemu-devel] [RFC PATCH] QEMU may write to system_memory before guest starts

[Eduardo Habkost]

My memory is failing here: do we still need to fix a bug where
there are unexpected writes to system_memory, or this is just a
request to include a mechanism to help us detect those cases in
the future?


On Tue, May 14, 2019 at 12:42:14PM +0300, Yury Kotov wrote:
> Ping ping
> 
> 17.04.2019, 15:46, "Yury Kotov" <address@hidden>:
> > Ping
> >
> > 04.04.2019, 13:01, "Yury Kotov" <address@hidden>:
> >>  I saw Catherine Ho's patch series and it seems ok to me, but in this RFC 
> >> I asked
> >>  about a way how to detect other writes which may not be covered by 
> >> particular
> >>  fixes.
> >>  Perhaps this is excessive caution...
> >>
> >>  Regards,
> >>  Yury
> >>
> >>  04.04.2019, 12:52, "Dr. David Alan Gilbert" <address@hidden>:
> >>>   * Юрий Котов (address@hidden) wrote:
> >>>>    Ping
> >>>
> >>>   Is this fixed by Catherine Ho's patch series?
> >>>
> >>>   Dave
> >>>
> >>>>    21.03.2019, 19:27, "Yury Kotov" <address@hidden>:
> >>>>    > Hi,
> >>>>    >
> >>>>    > 19.03.2019, 14:52, "Dr. David Alan Gilbert" <address@hidden>:
> >>>>    >>  * Peter Maydell (address@hidden) wrote:
> >>>>    >>>   On Tue, 19 Mar 2019 at 11:03, Dr. David Alan Gilbert
> >>>>    >>>   <address@hidden> wrote:
> >>>>    >>>   >
> >>>>    >>>   > * Peter Maydell (address@hidden) wrote:
> >>>>    >>>   > > I didn't think migration distinguished between "main memory"
> >>>>    >>>   > > and any other kind of RAMBlock-backed memory ?
> >>>>    >>>   >
> >>>>    >>>   > In Yury's case there's a distinction between RAMBlock's that 
> >>>> are mapped
> >>>>    >>>   > with RAM_SHARED (which normally ends up as MAP_SHARED) and 
> >>>> all others.
> >>>>    >>>   > You can set that for main memory by using -numa to specify a 
> >>>> memdev
> >>>>    >>>   > that's backed by a file and has the share=on property.
> >>>>    >>>   >
> >>>>    >>>   > On x86 the ROMs end up as separate RAMBlock's that aren't 
> >>>> affected
> >>>>    >>>   > by that -numa/share=on - so they don't fight Yury's trick.
> >>>>    >>>
> >>>>    >>>   You can use the generic loader on x86 to load an ELF file
> >>>>    >>>   into RAM if you want, which would I think also trigger this.
> >>>>    >>
> >>>>    >>  OK, although that doesn't worry me too much - since in the 
> >>>> majority
> >>>>    >>  of cases Yury's trick still works well.
> >>>>    >>
> >>>>    >>  I wonder if there's a way to make Yury's code to detect these 
> >>>> cases
> >>>>    >>  and not allow the feature; the best thing for the moment would 
> >>>> seem to
> >>>>    >>  be to skip the aarch test that uses elf loading.
> >>>>    >
> >>>>    > Currently, I've no idea how to detect such cases, but there is an 
> >>>> ability to
> >>>>    > detect memory corruption. I want to update the RFC patch to let 
> >>>> user to map some
> >>>>    > memory regions as readonly until incoming migration start.
> >>>>    >
> >>>>    > E.g.
> >>>>    > 1) If x-ignore-shared is enabled in command line or memory region 
> >>>> is marked
> >>>>    >    (something like ',readonly=on'),
> >>>>    > 2) Memory region is shared (,share=on),
> >>>>    > 3) And qemu is started with '-incoming' option
> >>>>    >
> >>>>    > Then map such regions as readonly until incoming migration finished.
> >>>>    > Thus, the patch will be able to detect memory corruption and will 
> >>>> not affect
> >>>>    > normal cases.
> >>>>    >
> >>>>    > How do you think, is it needed?
> >>>>    >
> >>>>    > I already have a cleaner version of the RFC patch, but I'm not sure 
> >>>> about 1).
> >>>>    > Which way is better: enable capability in command line, add a new 
> >>>> option for
> >>>>    > memory-backend or something else.
> >>>>    >
> >>>>    >>  Dave
> >>>>    >>
> >>>>    >>>   thanks
> >>>>    >>>   -- PMM
> >>>>    >>  --
> >>>>    >>  Dr. David Alan Gilbert / address@hidden / Manchester, UK
> >>>>    >
> >>>>    > Regards,
> >>>>    > Yury
> >>>   --
> >>>   Dr. David Alan Gilbert / address@hidden / Manchester, UK

-- 
Eduardo