Am 19.06.2012 20:02, schrieb Blue Swirl:
On Tue, Jun 19, 2012 at 5:53 AM, Stefan Weil <s...@weilnetz.de> wrote:
Am 19.06.2012 04:31, schrieb Scott Wood:

If tci_out_label is called in the context of tcg_gen_code_search_pc, we
could be overwriting an already patched relocation with zero -- and not
repatch it because the set_label is past search_pc, causing a QEMU crash
when it tries to branch to a zero label.

Not writing anything to the relocation area seems to be in line with what
other backends do from the couple I looked at (x86, ppc).


Thanks, this might fix a crash which I have seen from time to time.
I'll run tests as soon as possible.

Tested-by: Stefan Weil <s...@weilnetz.de>

The patch fixes my test scenario with a guest booting a Debian ARM kernel
on a x86_64 host (Linux or W64).

The following command crashes while the ARM Linux guest is booting
(shortly after "Freeing init memory") with SIGSEGV caused by tb_ptr == 2:

$ qemu-system-arm -M versatilepb -m 192 \
    -kernel vmlinuz-3.2.0-2-versatile -initrd initrd.img-3.2.0-2-versatile

With the patch applied, the ARM Linux boots correctly.

Blue, maybe that test also works with a SPARC host.
I have copied the ARM kernel and initrd to http://qemu.weilnetz.de/arm/.

Regards,
Stefan W.


Could you please also look at the other backends?

I saw from git history that ppc once had the same bug.
The sparc backend (and maybe others) might still have it.

Confirmed for Sparc.


Regards,
Stefan W.



Signed-off-by: Scott Wood<scottw...@freescale.com>
---
 tcg/tci/tcg-target.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/tcg/tci/tcg-target.c b/tcg/tci/tcg-target.c
index 453f187..3c6b0f5 100644
--- a/tcg/tci/tcg-target.c
+++ b/tcg/tci/tcg-target.c
@@ -487,7 +487,7 @@ static void tci_out_label(TCGContext *s, TCGArg arg)
         assert(label->u.value);
     } else {
tcg_out_reloc(s, s->code_ptr, sizeof(tcg_target_ulong), arg, 0);
-        tcg_out_i(s, 0);
+        s->code_ptr += sizeof(tcg_target_ulong);

I like this fix. Other similar fixes rewrote the fixed part of the
opcode and not the label, but the fixed part may cross byte
boundaries.

     }
 }





Reply via email to