<div dir="ltr">Please see comments below,<div><br></div><div>Cheers,</div><div>Alex</div><div class="gmail_extra"><br><br><div class="gmail_quote">On Mon, Apr 15, 2013 at 1:49 PM, Iorga, Cristian <span dir="ltr"><<a href="mailto:cristian.iorga@intel.com" target="_blank">cristian.iorga@intel.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Hi Alex,<br>
<br>
See my comments below.<br>
<br>
Regards,<br>
Cristian<br>
<br>
From: Damian, Alexandru [mailto:<a href="mailto:alexandru.damian@intel.com">alexandru.damian@intel.com</a>]<br>
Sent: Monday, April 15, 2013 3:31 PM<br>
To: Iorga, Cristian<br>
Cc: <a href="mailto:richard.purdie@linuxfoundation.org">richard.purdie@linuxfoundation.org</a>; <a href="mailto:tom.zanussi@linux.intel.com">tom.zanussi@linux.intel.com</a>; Burton, Ross; <a href="mailto:paul.eggleton@linux.intel.com">paul.eggleton@linux.intel.com</a>; <a href="mailto:openembedded-core@lists.openembedded.org">openembedded-core@lists.openembedded.org</a><br>
Subject: Re: [OE-core] [PATCH] qemu: use default qemu CPU definition<br>
<div class="im"><br>
Hi Cristian,<br>
<br>
Please do post comments on patch on the mailing list.<br>
<br>
Regarding the message itself, quoted below, I am addressing your specific concerns.<br>
- There is no performance impact between "qemu64" and "host" CPU options. The only thing changing is the level of private instructions that the kernel will attempt - which should be clearly restricted to "qemu64" and not "host" since we build Yocto for the "qemu" machine.<br>
</div> Agree.<br>
<div class="im">- I agree that there is an underlying problem in the kernel, but that is not our primary concern, see no 1. It might be our concern if we'd build for a specific BSP which showed the problem, but as it stands, it's KVM folk's concern. Funny enough, they know about the problem (I tracked on their mailing list), but they don't have a fix since it runs as it's designed right now - that is, the faulty behavior is a result of how it should work with the current design.<br>
</div> I don’t know the problems exactly in this case, but:<br>
<br>
- I suggested that the differences between how this issue is reproducible or not of different host kernels is strange, and I fail to see the logical connection between your patch and the aforementioned behaviors. Can you please provide an explanation?<br>
<br>
- It is also strange that this issue is reproducible only on qemux86-64 and not on qemux86 (qemux86 also exports CPU of the VM as host CPU), which suggests a different root cause.<br></blockquote><div><br></div><div>
<br></div><div style>The root cause is this: on specific CPU hosts, the kernel is trying to execute a "read model specific register" instruction. This is not handled by the qemu, and it's executed directly by host in KVM emulation mode. This triggers a GP0 fault, as designed, since it's not executed in ring 0 - it's executed in ring 3 since KVM is executing in userspace. So the guest now executes a GP0 handler which stops the guest machine.</div>
<div style><br></div><div style>The fix is actually preventing execution of the RDMSR/WRMSR instructions in KVM mode since the guest kernel know knows it's under emulation. </div><div> </div><div style>This is what I mean that the system actually works as designed.</div>
<div style><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div class="im"><br>
I still recommend merging this patch since it fixes the problem with the right fix.<br>
<br>
</div>- That’s a fix, but is it the right one, given my previous observations?<br></blockquote><div><br></div><div style>Yes, it is. </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
Also, may I remind you that this a regression?<br>
This issue was not present with the same kernel, on the same distro and it is not present on qemux86 VMs.<br>
<div class="im"><br>
Cheers,<br>
Alex<br>
<br>
Original message:<br>
<br>
Hi Alex,<br>
<br>
I might be wrong, and I need more time to investigate, but I don't agree with the fix you proposed for this issue.<br>
<br>
<br>
<br>
What you propose is just a step back from how qemu should optimally work, meaning exporting "-cpu host".<br>
<br>
<br>
<br>
That's one thing.<br>
<br>
<br>
<br>
Another thing is the described issue seems to be specific to some kernels on the host.<br>
<br>
For example, I checked with LaurP and AlexG and thes issue is:<br>
<br>
- reproducible on Ubuntu 12.10 with kernel 3.5.0.27 and Fedora Core 18 with kernel 3.8.5.<br>
<br>
- not reproducible with Ubuntu 12.04 with kernel 3.2.x (need to check the exact details with LaurP).<br>
<br>
<br>
<br>
This issue needs more investigation.<br>
<br>
<br>
<br>
Regards,<br>
<br>
Cristian<br>
<br>
</div><div class="im">On Mon, Apr 15, 2013 at 9:58 AM, Iorga, Cristian <<a href="mailto:cristian.iorga@intel.com">cristian.iorga@intel.com</a><mailto:<a href="mailto:cristian.iorga@intel.com">cristian.iorga@intel.com</a>>> wrote:<br>
Hi Alex,<br>
<br>
See my comments regarding this patch at:<br>
<a href="https://bugzilla.yoctoproject.org/show_bug.cgi?id=3908" target="_blank">https://bugzilla.yoctoproject.org/show_bug.cgi?id=3908</a><br>
<br>
Regards,<br>
Cristian<br>
<br>
-----Original Message-----<br>
</div><div class="im">From: <a href="mailto:openembedded-core-bounces@lists.openembedded.org">openembedded-core-bounces@lists.openembedded.org</a><mailto:<a href="mailto:openembedded-core-bounces@lists.openembedded.org">openembedded-core-bounces@lists.openembedded.org</a>> [mailto:<a href="mailto:openembedded-core-bounces@lists.openembedded.org">openembedded-core-bounces@lists.openembedded.org</a><mailto:<a href="mailto:openembedded-core-bounces@lists.openembedded.org">openembedded-core-bounces@lists.openembedded.org</a>>] On Behalf Of Alex DAMIAN<br>
Sent: Tuesday, April 09, 2013 5:47 PM<br>
</div><div class="im">To: <a href="mailto:richard.purdie@linuxfoundation.org">richard.purdie@linuxfoundation.org</a><mailto:<a href="mailto:richard.purdie@linuxfoundation.org">richard.purdie@linuxfoundation.org</a>>; <a href="mailto:tom.zanussi@linux.intel.com">tom.zanussi@linux.intel.com</a><mailto:<a href="mailto:tom.zanussi@linux.intel.com">tom.zanussi@linux.intel.com</a>>; Burton, Ross; <a href="mailto:paul.eggleton@linux.intel.com">paul.eggleton@linux.intel.com</a><mailto:<a href="mailto:paul.eggleton@linux.intel.com">paul.eggleton@linux.intel.com</a>>; <a href="mailto:openembedded-core@lists.openembedded.org">openembedded-core@lists.openembedded.org</a><mailto:<a href="mailto:openembedded-core@lists.openembedded.org">openembedded-core@lists.openembedded.org</a>><br>
Cc: Damian, Alexandru<br>
Subject: [OE-core] [PATCH] qemu: use default qemu CPU definition<br>
<br>
</div>From: Alexandru DAMIAN <<a href="mailto:alexandru.damian@intel.com">alexandru.damian@intel.com</a><mailto:<a href="mailto:alexandru.damian@intel.com">alexandru.damian@intel.com</a>>><br>
<div class="im"><br>
This fixes guest crashes when trying to use restricted instructions based on CPU description. Using default qemu CPU descriptions restrict faults on using restricted model-specific registers.<br>
<br>
[YOCTO #3908]<br>
<br>
</div>Signed-off-by: Alexandru DAMIAN <<a href="mailto:alexandru.damian@intel.com">alexandru.damian@intel.com</a><mailto:<a href="mailto:alexandru.damian@intel.com">alexandru.damian@intel.com</a>>><br>
<div><div class="h5">---<br>
scripts/runqemu | 2 +-<br>
scripts/runqemu-internal | 6 ++++--<br>
2 files changed, 5 insertions(+), 3 deletions(-)<br>
<br>
diff --git a/scripts/runqemu b/scripts/runqemu index 8ed1226..43d7600 100755<br>
--- a/scripts/runqemu<br>
+++ b/scripts/runqemu<br>
@@ -261,7 +261,7 @@ if [ "x$KVM_ENABLED" = "xyes" ]; then<br>
exit 1;<br>
fi<br>
if [ -w /dev/kvm -a -r /dev/kvm ]; then<br>
- SCRIPT_QEMU_OPT="$SCRIPT_QEMU_OPT -enable-kvm -cpu host"<br>
+ SCRIPT_QEMU_OPT="$SCRIPT_QEMU_OPT -enable-kvm "<br>
KVM_ACTIVE="yes"<br>
else<br>
echo "You have no rights on /dev/kvm."<br>
diff --git a/scripts/runqemu-internal b/scripts/runqemu-internal index e8191d1..a7813e5 100755<br>
--- a/scripts/runqemu-internal<br>
+++ b/scripts/runqemu-internal<br>
@@ -323,10 +323,11 @@ fi<br>
<br>
if [ "$MACHINE" = "qemux86" ]; then<br>
QEMU=qemu-system-i386<br>
+ QEMU_CPU_OPTIONS="-cpu qemu32"<br>
QEMU_UI_OPTIONS="$QEMU_UI_OPTIONS -vga vmware"<br>
if [ "$FSTYPE" = "ext2" -o "$FSTYPE" = "ext3" -o "$FSTYPE" = "btrfs" ]; then<br>
KERNCMDLINE="vga=0 uvesafb.mode_option=640x480-32 root=$DROOT rw mem=$QEMU_MEMORY $KERNEL_NETWORK_CMD"<br>
- QEMUOPTIONS="$QEMU_NETWORK_CMD $ROOTFS_OPTIONS $QEMU_UI_OPTIONS"<br>
+ QEMUOPTIONS="$QEMU_NETWORK_CMD $ROOTFS_OPTIONS $QEMU_CPU_OPTIONS $QEMU_UI_OPTIONS"<br>
fi<br>
if [ "$FSTYPE" = "nfs" ]; then<br>
if [ "$NFS_SERVER" = "192.168.7.1" -a ! -d "$NFS_DIR" ]; then @@ -347,10 +348,11 @@ fi<br>
<br>
if [ "$MACHINE" = "qemux86-64" ]; then<br>
QEMU=qemu-system-x86_64<br>
+ QEMU_CPU_OPTIONS="-cpu qemu64"<br>
QEMU_UI_OPTIONS="$QEMU_UI_OPTIONS -vga vmware"<br>
if [ "$FSTYPE" = "ext3" -o "$FSTYPE" = "btrfs" ]; then<br>
KERNCMDLINE="vga=0 root=$DROOT rw mem=$QEMU_MEMORY $KERNEL_NETWORK_CMD"<br>
- QEMUOPTIONS="$QEMU_NETWORK_CMD $ROOTFS_OPTIONS $QEMU_UI_OPTIONS"<br>
+ QEMUOPTIONS="$QEMU_NETWORK_CMD $ROOTFS_OPTIONS $QEMU_CPU_OPTIONS $QEMU_UI_OPTIONS"<br>
fi<br>
if [ "$FSTYPE" = "nfs" ]; then<br>
if [ "x$ROOTFS" = "x" ]; then<br>
--<br>
1.7.10.4<br>
<br>
_______________________________________________<br>
Openembedded-core mailing list<br>
</div></div><a href="mailto:Openembedded-core@lists.openembedded.org">Openembedded-core@lists.openembedded.org</a><mailto:<a href="mailto:Openembedded-core@lists.openembedded.org">Openembedded-core@lists.openembedded.org</a>><br>
<a href="http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core" target="_blank">http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core</a><br>
<br>
</blockquote></div><br></div></div>