Re: pstore does not work under xen

[Boris Ostrovsky]

On 9/23/19 6:59 PM, Kees Cook wrote:
> On Mon, Sep 23, 2019 at 03:42:27PM +0000, James Dingwall wrote:
>> On Thu, Sep 19, 2019 at 12:37:40PM -0400, Boris Ostrovsky wrote:
>>> On 9/19/19 12:14 PM, James Dingwall wrote:
>>>> On Thu, Sep 19, 2019 at 03:51:33PM +0000, Luck, Tony wrote:
>>>>>> I have been investigating a regression in our environment where pstore 
>>>>>> (efi-pstore specifically but I suspect this would affect all 
>>>>>> implementations) no longer works after upgrading from a 4.4 to 5.0 
>>>>>> kernel when running under xen.  (This is an Ubuntu kernel but I don't 
>>>>>> think there are patches which affect this area.)
>>>>> I don't have any answer for this ... but want to throw out the idea that
>>>>> VMM systems could provide some hypercalls to guests to save/return
>>>>> some blob of memory (perhaps the "save" triggers automagically if the
>>>>> guest crashes?).
>>>>>
>>>>> That would provide a much better pstore back end than relying on emulation
>>>>> of EFI persistent variables (which have severe contraints on size, and don't
>>>>> support some pstore modes because you can't dynamically update EFI variables
>>>>> hundreds of times per second).
>>>>>
>>>> For clarification this is a dom0 crash rather than an HVM guest with EFI.  I
>>>> should probably have also mentioned the xen verion has changed from 4.8.4 to
>>>> 4.11.2 in case its behaviour on detection of crashed domain has changed.
>>>>
>>>> (For capturing guest crashes we have enabled xenconsole logging so the
>>>> hvc0 log is available in dom0.)
>>>
>>> Do you only see this difference between 4.4 and 5.0 when you crash via
>>> sysrq?
>>>
>>> Because that's where things changed. On 4.4 we seem to be forcing an
>>> oops, which eventually calls kmsg_dump() and then panic. On 5.0 we call
>>> panic() directly from sysrq handler. And because Xen's panic notifier
>>> doesn't return we never get a chance to call kmsg_dump().
>>>
>> Ok, I see that change in 8341f2f222d729688014ce8306727fdb9798d37e.  I 
>> hadn't tested it any other way before.  Using the null pointer 
>> de-reference module code at [1] a pstore record is generated as expected 
>> when the module is loaded (panic_on_oops=1).
> This change looks correct -- it just gets us directly to the panic()
> state instead of exercising the various exception handlers.
>
>> I have also tested swapping the kmsg_dump() / 
>> atomic_notifier_call_chain() around in panic.c and this also results in 
>> a pstore record being created with sysrq-c.  I don't know if that would 
>> be an acceptable solution though since it may break behaviour that other 
>> things depend on.
> I don't think reordering these is a good idea: as the comments say,
> there might be work done in the notifier chain that kmsg_dump() will
> want to capture (e.g. the KASLR base offset).
>
> The situation seems to be that notifier callbacks must return -- I think
> Xen needs fixing here.
>

I only had one quick sanity test with a PV guest so this needs more
testing. James, can you give it a try?


diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index 750f46ad018a..d88f118028b4 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -269,16 +269,17 @@ void xen_reboot(int reason)
ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ BUG();
Â}
Â
+static int reboot_reason = SHUTDOWN_reboot;
Âvoid xen_emergency_restart(void)
Â{
-ÂÂÂÂÂÂ xen_reboot(SHUTDOWN_reboot);
+ÂÂÂÂÂÂ xen_reboot(reboot_reason);
Â}
Â
Âstatic int
Âxen_panic_event(struct notifier_block *this, unsigned long event, void
*ptr)
Â{
ÂÂÂÂÂÂÂ if (!kexec_crash_loaded())
-ÂÂÂÂÂÂÂÂÂÂÂÂÂÂ xen_reboot(SHUTDOWN_crash);
+ÂÂÂÂÂÂÂÂÂÂÂÂÂÂ reboot_reason = SHUTDOWN_crash;
ÂÂÂÂÂÂÂ return NOTIFY_DONE;
Â}
Â
@@ -290,6 +291,8 @@ static struct notifier_block xen_panic_block = {
Âint xen_panic_handler_init(void)
Â{
ÂÂÂÂÂÂÂ atomic_notifier_chain_register(&panic_notifier_list,
&xen_panic_block);
+ÂÂÂÂÂÂ if (panic_timeout == 0)
+ÂÂÂÂÂÂÂÂÂÂÂÂÂÂ set_arch_panic_timeout(-1, CONFIG_PANIC_TIMEOUT);
ÂÂÂÂÂÂÂ return 0;
Â}