[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] HVM domains crash after upgrade from XEN 4.5.1 to 4.5.2

On 14/11/2015 00:16, Atom2 wrote:
Am 13.11.15 um 11:09 schrieb Andrew Cooper:
On 13/11/15 07:25, Jan Beulich wrote:
On 13.11.15 at 00:00, <ariel.atom2@xxxxxxxxxx> wrote:
Am 12.11.15 um 17:43 schrieb Andrew Cooper:
On 12/11/15 14:29, Atom2 wrote:
Hi Andrew,
thanks for your reply. Answers are inline further down.

Am 12.11.15 um 14:01 schrieb Andrew Cooper:
On 12/11/15 12:52, Jan Beulich wrote:
On 12.11.15 at 02:08, <ariel.atom2@xxxxxxxxxx> wrote:
After the upgrade HVM domUs appear to no longer work - regardless
of the
dom0 kernel (tested with both 3.18.9 and 4.1.7 as the dom0 kernel); PV
domUs, however, work just fine as before on both dom0 kernels.

xl dmesg shows the following information after the first crashed HVM
domU which is started as part of the machine booting up:
(XEN) Failed vm entry (exit reason 0x80000021) caused by invalid guest
state (0).
(XEN) ************* VMCS Area **************
(XEN) *** Guest State ***
(XEN) CR0: actual=0x0000000000000039, shadow=0x0000000000000011,
(XEN) CR4: actual=0x0000000000002050, shadow=0x0000000000000000,
(XEN) CR3: actual=0x0000000000800000, target_count=0
(XEN)      target0=0000000000000000, target1=0000000000000000
(XEN)      target2=0000000000000000, target3=0000000000000000
(XEN) RSP = 0x0000000000006fdc (0x0000000000006fdc)  RIP =
0x0000000100000000 (0x0000000100000000)
Other than RIP looking odd for a guest still in non-paged protected
mode I can't seem to spot anything wrong with guest state.
odd? That will be the source of the failure.

Out of long mode, the upper 32bit of %rip should all be zero, and it
should not be possible to set any of them.

I suspect that the guest has exited for emulation, and there has been a
bad update to %rip.  The alternative (which I hope is not the case) is
that there is a hardware errata which allows the guest to accidentally
get it self into this condition.

Are you able to rerun with a debug build of the hypervisor?
Another question is whether prior to enabling the debug USE flag it
might make sense to re-compile with gcc-4.8.5 (please see my previous
list reply) to rule out any compiler related issues. Jan, Andrew -
what are your thoughts?
First of all, check whether the compiler makes a difference on 4.5.2
Hi Andrew,
I changed the compiler and there was no change to the better:
Unfortunately the HVM domU is still crashing with a similar error
message as soon as it is being started.
If both compiles result in a guest crashing in that manner, test a debug
Xen to see if any assertions/errors are encountered just before the
guest crashes.

As the compiler did not make any difference, I enabled the debug USE
flag, re-compiled (using gcc-4.9.3), and rebooted using a serial console
to capture output. Unfortunately I did not get very far and things
become even stranger: This time the system did not even finnish the boot
process, but rather hard-stopped pretty early with a message reading
"Panic on CPU 3: DOUBLE FAULT -- system shutdown". The captured logfile
is attached as "serial log.txt".

As this happened immediately after the CPU microcode update, I thought
there might be a connection and disabled the microcode update. After the
next reboot it seemed as if the boot process got a bit further as
evidenced by a few more lines in the log file (those between lines 136
and 197 in the second log file named "serial log no ucode.txt"), but in
the end it finnished off with an identical error message (only the CPU #
was different this time, but that number seems to change between boots

I hope that makes some sense to you.
Not really, other than now even more suspecting bad hardware or
something fundamentally wrong with your build. Did you retry with
a freshly built 4.5.1? Could you alternatively try with a known good
build of 4.5.2 (e.g. from osstest)?
many thanks again for your help.
Agreed.  Double faults indicate that the exception handing entry points
are not set up in an appropriate state.  Something is definitely wrong
with either the compiled binary or the hardware.
The hardware (it's a SandyBridge XEON processor with ECC RAM and Enterprise SATA disks) has worked for almost two years together with XEN and other than this issue there's also currently nothing strange (i.e. if I boot with a standard linux kernel, the system boots and works without any issues and is very stable and there are also no strange messages in /var/log/messages).
Several questions and lines of investigation:

Is this straight Xen 4.5.1 and 2, or do Gentoo have their own patches on
By the looks of it there are only security patches, but no gentoo specific patches.
On repeated attempts, are the details of the double fault identical
(other than the cpu), or does it move around (i.e. always do_IRQ+0x15)
It always seems to be do_IRQ+0x15 (I have made a number of boots), and more specifically it was always do_IRQ+0x15/0x64c. Timings varied, the CPU differed between boots and also the rax, rbx, rcx, rdx, rsi, rdi, rbp, rsp, and r8 values were diffent as were those for r15 and cr2 and the values next to valid stack range. I don't know whether that's of relevance, but I thought I'd mention it after a quick analysis of two serial logs.
Can you boot with console_timestamps=boot on the command line in the
future.  This will put Linux-sytle timestamps on log messages.
Done - see two the attached serial log files mentioned above.
Can you also compile in the attached patch? I haven't quite got it
suitable for inclusion upstream yet, but it will also dump the
instruction stream under the fault.
I have added the patch, but it seems not to trigger - at least the text that is in the patch does not show in the serial console output; it is however in the (uncompressed) xen.gz file as a quick grep for texts "Xen code around " and " [fault on access]" confirmed (grep said: binary file matches).
Finally, can you disassemble the xen-syms which results from the debug
build and paste the start of do_IRQ.  (i.e. `gdb xen-syms` and "disass
Please see the third file named "do_IRQ".

Furthermore I have managed to get the system to again boot up (a HVM domU, however, still crashes): If I turn off the debug USE flag and just add symbols to the XEN binary. It seems that the debug USE flag is probably not what I was expecting. On https://wiki.gentoo.org/wiki/Project:Quality_Assurance/Backtraces it describes the purpose of the debug USE flag as follows (emphasis by me):

==== start excerpt from web page ====
Some ebuilds provide a debug USE flag. While some mistakenly use it to provide debug information and play with compiler flags when it is enabled, that is not its purpose.

_If you're trying to debug a reproduceable crash, you want to leave this USE flag alone, as it'll be building a different source than what you had before._ It is more efficient to get first a backtrace without changing the code, by simply emitting symbol information, and just afterward enable debug features to track the issue further down.

Debug features that are enabled by the USE flag include assertions, debug logs on screen, debug files, leak detection and extra-safe operations (such as scrubbing memory before use). Some of them might be taxing, especially for complex software or software where performance is an important issue.

For these reasons, please exercise caution when enabling the debug USE flag, and only consider it a last-chance card.
==== end excerpt from web page ====

Now _without_ the debug USE flag, but with debug information in the binary (I used splitdebug), all is back to where the problem started off (i.e. the system boots without issues until such time it starts a HVM domU which then crashes; PV domUs are working). I have attached the latest "xl dmesg" output with the timing information included.

I hope any of this makes sense to you.

Again many thanks and best regards

Right - it would appear that the USE flag is definitely not what you wanted, and causes bad compilation for Xen. The do_IRQ disassembly you sent is a the result of disassembling a whole block of zeroes. Sorry for leading you on a goose chase - the double faults will be the product of bad compilation, rather than anything to do with your specific problem.

However, the final log you sent (dmesg) is using a debug Xen, which is what I was attempting to get you to do originally.

We still observe that the VM ends up in 32bit non-paged mode but with an RIP with bit 32 set, which is an invalid state to be in. However, there was nothing particularly interesting in the extra log information.

Please can you rerun with "hvm_debug=0xc3f", which will cause far more logging to occur to the console while the HVM guest is running. That might show some hints.

Also, the fact that this occurs just after starting SeaBIOS is interesting. As you have switched versions of Xen, you have also switched hvmloader, which contains the SeaBIOS binary embedded in it. Would you be able to compile both 4.5.1 and 4.5.2 and switch the hvmloader binaries in use. It would be very interesting to see whether the failure is caused by the hvmloader binary or the hypervisor. (With `xl`, you can use firmware_override="/full/path/to/firmware" to override the default hvmloader).


Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.