Re: [Qemu-devel] [Qemu-ppc] standard test image not booting with qemu-system-ppc

[Alexander Graf]

On 03.02.2014, at 23:58, Alexander Graf <address@hidden> wrote:

> 
> On 03.02.2014, at 22:13, Alexander Graf <address@hidden> wrote:
> 
>> 
>> On 03.02.2014, at 21:44, Mark Cave-Ayland <address@hidden> wrote:
>> 
>>> On 03/02/14 20:02, Nitin Srivastava wrote:
>>> 
>>>> Hi ,
>>>> I downloaded the latest qemu source from git and compiled it on my
>>>> centos 6.5 machine.
>>>> According to the following e-mail from this mailing list I tried the
>>>> following
>>>> _http://comments.gmane.org/gmane.comp.emulators.qemu/181171_
>>>> Test case: download the squeeze standard image from
>>>> _http://people.debian.org/~aurel32/qemu/powerpc/_
>>>> <http://people.debian.org/%7Eaurel32/qemu/powerpc/>
>>>> and run with
>>>> qemu-system-ppc -hda debian_squeeze_powerpc_standard.qcow2
>>>> but I get following error
>>>> VNC server running on `::1:5900'
>>>>>> =============================================================
>>>>>> OpenBIOS 1.1 [Oct 2 2013 22:57]
>>>>>> Configuration device id QEMU version 1 machine id 2
>>>>>> CPUs: 1
>>>>>> Memory: 128M
>>>>>> UUID: 00000000-0000-0000-0000-000000000000
>>>>>> CPU type PowerPC,750
>>>>>> Not a bootable ELF image
>>>> qemu: terminating on signal 2
>>>> address@hidden:~:117#
>>>> also please note that my qemu-system-ppc is latest, as its built from
>>>> source.
>>>> address@hidden:~:117#qemu-system-ppc -version
>>>> QEMU emulator version 1.7.50, Copyright (c) 2003-2008 Fabrice Bellard
>>>> address@hidden:~:118#
>>>> Please help.
>>>> Regds.
>>>> Nitin
>>> 
>>> Hi Nitin,
>>> 
>>> Having just updated to git master, I now see this issue too. A quick 
>>> session with git bisect shows the culprit is this commit:
>>> 
>>> 
>>> address@hidden:~/src/qemu/git/qemu$ git bisect bad
>>> 360e607b88a23d378f6efaa769c76d26f538234d is the first bad commit
>>> commit 360e607b88a23d378f6efaa769c76d26f538234d
>>> Author: Stefano Stabellini <address@hidden>
>>> Date:   Thu Jan 30 12:46:05 2014 +0000
>>> 
>>>  address_space_translate: do not cross page boundaries
>>> 
>>> The following commit:
>>> 
>>> commit 149f54b53b7666a3facd45e86eece60ce7d3b114
>>> Author: Paolo Bonzini <address@hidden>
>>> Date:   Fri May 24 12:59:37 2013 +0200
>>> 
>>>  memory: add address_space_translate
>>> 
>>> breaks Xen support in QEMU, in particular the Xen mapcache. The effect
>>> is that one Windows XP installation out of ten would end up with BSOD.
>>> 
>>> The reason is that after this commit l in address_space_rw can span a
>>> page boundary, however qemu_get_ram_ptr still calls xen_map_cache asking
>>> to map a single page (if block->offset == 0).
>>> 
>>> Fix the issue by reverting to the previous behaviour: do not return a
>>> length from address_space_translate_internal that can span a page
>>> boundary.
>>> 
>>> Also in address_space_translate do not ignore the length returned by
>>> address_space_translate_internal.
>>> 
>>> This patch should be backported to QEMU 1.6.x.
>>> 
>>> Signed-off-by: Stefano Stabellini <address@hidden>
>>> Signed-off-by: Anthony Perard <address@hidden>
>>> Tested-by: Paolo Bonzini <address@hidden>
>>> Acked-by: Paolo Bonzini <address@hidden>
>>> Cc: address@hidden
>>> 
>>> 
>>> Stefano/Alex, is there any reason why this would break qemu-system-ppc?
>> 
>> Ugh, sorry Nitin, I should have read the email to the end.
>> 
>> The image does work for me with -nographic, so I'd assume it's something 
>> about the frame buffer map going wrong? We do successfully run the guest:
>> 
>> address@hidden:/home/agraf/release/qemu> ./ppc-softmmu/qemu-system-ppc -vnc 
>> :8 -snapshot -hda /dev/shm/debian_squeeze_powerpc_standard.qcow2 -serial 
>> mon:stdio
>> 
>>>> =============================================================
>>>> OpenBIOS 1.1 [Oct 2 2013 22:57]
>>>> Configuration device id QEMU version 1 machine id 2
>>>> CPUs: 1
>>>> Memory: 128M
>>>> UUID: 00000000-0000-0000-0000-000000000000
>>>> CPU type PowerPC,750
>>>> Not a bootable ELF image
>> QEMU 1.7.50 monitor - type 'help' for more information
>> (qemu) x /i $pc
>> 0xfff0c2e0:  b       0xfff0c2ec
>> (qemu) x /i $pc
>> 0xfff25dd4:  lwz     r0,4(r11)
>> (qemu) x /i $pc
>> 0xfff1552c:  lhz     r9,0(r4)
>> (qemu) x /i $pc
>> 0xfff0a868:  lwz     r0,20(r1)
>> (qemu) x /i $pc
>> 0xfff25d3c:  stw     r31,-4(r11)
>> (qemu) x /i $pc
>> 0xfff0aeb8:  mr      r31,r3
>> (qemu) x /i $pc
>> 0xfff0b050:  mr      r31,r3
>> (qemu) x /i $pc
>> 0xfff0a6cc:  lis     r9,-5
>> (qemu) x /i $pc
>> 0xc0252a5c:  mr      r3,r31
>> (qemu) x /i $pc
>> 0xc003e274:  lwz     r9,68(r30)
>> 
>> so it's really only the VGA output that's broken.
> 
> A simple git revert of Stefanos patch makes VGA work again. The diff of "info 
> mtree" of a mac99 system with and without his patch is the following:
> 
> --- x 2014-02-03 23:57:20.000000000 +0100
> +++ y 2014-02-03 23:56:14.000000000 +0100
> @@ -30,6 +30,7 @@
> aliases
> pci-mmio
> 0000000000000000-00000000ffffffff (prio 0, RW): pci-mmio
> +  00000000000a0000-00000000000affff (prio 2, RW): alias vga.chain4 @vga.vram 
> 0000000000000000-000000000000ffff
>   00000000000a0000-00000000000bffff (prio 1, RW): vga-lowmem
>   0000000080000000-0000000080ffffff (prio 1, RW): vga.vram
>   0000000081000000-0000000081000fff (prio 1, RW): vga.mmio
> @@ -77,6 +78,8 @@
>   00000000000003d4-00000000000003d5 (prio 0, RW): vga
>   00000000000003da-00000000000003da (prio 0, RW): vga
>   0000000000000400-00000000000004ff (prio 1, RW): ne2000
> +vga.vram
> +0000000080000000-0000000080ffffff (prio 1, RW): vga.vram
> escc-bar
> 0000000000013000-000000000001303f (prio 0, RW): alias escc-bar @escc 
> 0000000000000000-000000000000003f
> escc
> 
> I'm still not quite sure what did cause the breakage.

Consider me still confused. The following patch makes it work again:

diff --git a/exec.c b/exec.c
index 9ad0a4b..dcc0fc1 100644
--- a/exec.c
+++ b/exec.c
@@ -351,7 +351,7 @@ MemoryRegion *address_space_translate(AddressSpace *as, 
hwaddr addr,
     hwaddr len = *plen;

     for (;;) {
-        section = address_space_translate_internal(as->dispatch, addr, &addr, 
&len, true);
+        section = address_space_translate_internal(as->dispatch, addr, &addr, 
plen, true);
         mr = section->mr;

         if (!mr->iommu_ops) {

The problem seems to be that 

But we do set *plen after the loop, so why would this change possibly change 
anything? This patch for example does _not_ make it work:

diff --git a/exec.c b/exec.c
index 9ad0a4b..a37c5f4 100644
--- a/exec.c
+++ b/exec.c
@@ -352,6 +352,7 @@ MemoryRegion *address_space_translate(AddressSpace *as, 
hwaddr addr,

     for (;;) {
         section = address_space_translate_internal(as->dispatch, addr, &addr, 
&len, true);
+        *plen = len;
         mr = section->mr;

         if (!mr->iommu_ops) {

That means that something relies on the incorrect behavior we did before to set 
*plen (end) = *plen (start) after the loop when !mr->iommu_ops. I wonder what 
that could be...


Alex