[vbox-dev] Problems with porting a VritualBox frontend from Windows to Linux
Klaus Espenlaub
klaus.espenlaub at oracle.com
Tue Mar 31 15:46:30 GMT 2015
Hi Rūdolf,
On 31.03.2015 16:52, Rūdolfs Bundulis wrote:
> HI Michal,
>
> > You could sprinkle a few 'Assert(env == cpu_single_env)'
> statements in the recompiler code and see if any of them hit.
>
> thanks for the info, I'll try that out. Yeah, debugging the debug
> build with gdb under Windows is kind of uncomfortable, under Linux it
> went much smoother. If Linux runs fine I'm not that desperate about
> fixing the Windows crash but then again - we had an idea that we could
> make something similar to the OpenGL service for Direct3D
> guest-to-host acceleration (of course if both guest and host are
> Windows with matching Direct3D version support), that should be more
> snappy than wine translating stuff into OpenGL. If we actually achieve
> that then I guess you could be interested into merging that in. Thanks
> again.
Replying off the list because I don't want to discuss this much in
public at this point - might attract people with the wrong intentions...
There's the VMSVGA2 device emulation (these days 2D+3D) contribution
which might do most of what you need. VMware specified the 3D interface
along Direct3D concepts (which means that e.g. Linux VMs need to get
their shaders translated), which possibly helps a great deal.
The catch is - we don't have a driver for Windows guests, as the
contribution focused on the raw device compatibility and the company
doing the funding wasn't interested at all in Windows (or even Linux -
but as the Linux VMSVGA2 drivers are open source they work nicely and
have been tested).
I would be very surprised if VMware's driver for Windows (should be
included in their tools installer) would work in VirtualBox. We never
tried, and probably won't try as we're afraid of legal issues, leave
alone that it'd be embarrassing to refer to the biggest competitor for
getting drivers. Are you interested in convincing our WDDM driver to do
2D/3D with VMSVGA2 device backing - either as a separate, standalone
driver or making our current one talk to both VBoxVGA and VMSVGA2?
Regards,
Klaus
>
> Best Regards,
> Rudolfs Bundulis
>
> 2015-03-31 17:40 GMT+03:00 Michal Necasek <michal.necasek at oracle.com
> <mailto:michal.necasek at oracle.com>>:
>
>
> Rudolfs,
>
> The recompiler crash you had there is not the same as what we just
> debugged but it is similar.
>
> The difference is that what we saw only affected 32-bit Windows. It
> only happened with release builds, not debug builds.
>
> The problem was a bug in the gcc compiler triggered by a large
> amount
> of local variables in a function. One or more registers saved on the
> stack got destroyed.
>
> In the recompiler code, generally it must hold that env ==
> cpu_single_env. If that is not true then 'env' probably got destroyed.
> Note that gdb will likely not display 'env' correctly.
>
> You could sprinkle a few 'Assert(env == cpu_single_env)'
> statements in
> the recompiler code and see if any of them hit.
>
>
> Regards,
> Michal
>
> On 3/31/2015 3:40 PM, Rūdolfs Bundulis wrote:
> > Hi Klaus,
> >
> > > Did you by any chance extend the CPUCTX struct or any other data
> > > structure which might be used by the recompiler? Just hours
> ago we
> > > tracked down a MingW gcc bug in the very same area which
> smells very
> > > similar...
> >
> > No, they only thing I've done is provided an IFramebuffer
> > implementation, I haven't touched any internals. I mean, so far
> all I
> > was doing was based on the SDK API, the init call now is the
> first thing
> > I'm using from the actual source tree. But something very bad
> could be
> > wrong with my implementation that could cause this, as I said, the
> > weirdest thing is that when running under Windows with a debug build
> > VirtualBox frontends never hit the recompiler but my code does.
> Now that
> > I have found a solution for Linux I'll do full testing and see if it
> > works fine - if it does , the issue is Windows only, if not, I
> should be
> > able to debug. Simply for me it's really hard to understand what is
> > going on in the recompiler code under Windows since I'm not familiar
> > with Qemu. I don't know if you had time to look at it - but I at
> least
> > posted the longjmp stack trace here
> >
> https://www.virtualbox.org/pipermail/vbox-dev/2015-February/012872.html,
> > is it similar to what you have found?
> >
> >
> >> "Regular" API clients don't need this, they can happily live
> with the
> >> implicit, delayed initialization which disables some functionality.
> >> That's why it makes no sense to talk about it in
> tstVBoxAPIXPCOM.cpp -
> >> it's happily working.
> >
> > Sorry,I did not quite understand - so the fact I was running fine
> > without it on Windows is because it is needed only for XPCOM and
> not COM
> > or was I just lucky?
> >
> >
> > Best Regards,
> > Rudolfs Bundulis
> >
> > 2015-03-31 16:16 GMT+03:00 Klaus Espenlaub
> <klaus.espenlaub at oracle.com <mailto:klaus.espenlaub at oracle.com>
> > <mailto:klaus.espenlaub at oracle.com
> <mailto:klaus.espenlaub at oracle.com>>>:
> >
> > Rūdolfs,
> >
> > On 31.03.2015 13:22, Rūdolfs Bundulis wrote:
> > > Hi,
> > >
> > > ok I found the root issue - my code did not call
> RTR3InitExe(argc,
> > > &argv, 0) since I built the application only from looking
> at the API
> > > interfaces and was not aware that I need to do this
> (actually maybe it
> > > makes sense to include a comment int the xpcom
> tstVBoxAPIXPCOM.cpp
> > > sample about this?).
> >
> > "Regular" API clients don't need this, they can happily live
> with the
> > implicit, delayed initialization which disables some
> functionality.
> > That's why it makes no sense to talk about it in
> tstVBoxAPIXPCOM.cpp -
> > it's happily working.
> >
> > Only API clients actually running VMs inside them absolutely
> need the
> > full initialization as the very first thing, otherwise they
> won't work.
> >
> > > Now after I've taken a closer look to the frontends
> > > I see that all them call the initialization function the
> first thing. I
> > > added it and my IFrambuffer implementation started to get
> invocations.
> >
> > Pretty sure that e.g. VBoxManage.cpp only does it because in
> uses quite
> > some runtime functionality, and doesn't want to run any risk
> (in the old
> > days it was far less "optional" to do the initialization).
> >
> > > Since my Windows code worked fine without it - should I
> add it to
> > > Windows as well? Any chance it is responsible for the
> weird recompiler
> > > longjmp crash I get? Thanks in advance.
> >
> > Did you by any chance extend the CPUCTX struct or any other data
> > structure which might be used by the recompiler? Just hours
> ago we
> > tracked down a MingW gcc bug in the very same area which
> smells very
> > similar...
> >
> > Regards
> > Klaus
> > >
> > > Best Regards,
> > > Rudolfs Bundulis
> > >
> > >
> > >
> > >2015-03-30 <tel:2015-03-30> <tel:2015-03-30
> <tel:2015-03-30>> 2:42 GMT+03:00 Rūdolfs Bundulis
> > <rudolfs.bundulis at gmail.com
> <mailto:rudolfs.bundulis at gmail.com>
> <mailto:rudolfs.bundulis at gmail.com
> <mailto:rudolfs.bundulis at gmail.com>>
> > > <mailto:rudolfs.bundulis at gmail.com
> <mailto:rudolfs.bundulis at gmail.com>
> > <mailto:rudolfs.bundulis at gmail.com
> <mailto:rudolfs.bundulis at gmail.com>>>>:
> > >
> > > Hi,
> > >
> > > I compiled debug version of 4.3.26 and tried my code
> against it -
> > > now the log has a much nicer assert message just
> before it ends:
> > >
> > > 00:00:55.619996 TM: GIP - u32Mode=1 (SyncTSC)
> u32UpdateHz=83
> > > 00:00:55.652313 TM: cTSCTicksPerSecond=0xa0861eba (2
> 693 144 250)
> > > fTSCVirtualized=true fTSCUseRealTSC=false
> > > 00:00:55.652322 TM: fMaybeUseOffsettedHostTSC=true
> > > TSCTiedToExecution=false TSCNotTiedToHalt=false
> > > 00:00:55.652373
> > > 00:00:55.652376 !!Assertion Failed!!
> > > 00:00:55.652378 Expression: <NULL>
> > > 00:00:55.652380 Location :
> > >
> > /home/rudolfs/VirtualBox-4.3.26/src/VBox/VMM/VMMR3/TM.cpp(576) int
> > > TMR3Init(PVM)
> > > 00:00:55.652508 Failed to create timer, u32Millies=10
> > > rc=VERR_NOT_SUPPORTED.
> > >
> > > I put a breakpoint in TM.cpp:576, and it seems that I
> am failing
> > > because of the code in RTTimerCreateEx():
> > >
> > > /*
> > > * We need the signal masks to be set correctly,
> which they
> > > won't be in
> > > * unobtrusive mode.
> > > */
> > > if (RTR3InitIsUnobtrusive())
> > > return VERR_NOT_SUPPORTED;
> > >
> > > What is unobtrusive mode and what signal masks do I
> need to set?
> > > I'll have a look at what is done in the frontends, but I
> > guess that
> > > you guys here can explain it in more detail.
> > >
> > > Best Reagrds,
> > > Rudolfs Bundulis
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.virtualbox.org/pipermail/vbox-dev/attachments/20150331/81dee4f2/attachment.html>
More information about the vbox-dev
mailing list