Linux Virtualization - Apr 2019 - [PATCH] x86/paravirt: Guard against invalid cpu # in pv_vcpu_is_preempted()

On 03/25/2019 12:40 PM, Juergen Gross wrote:
> On 25/03/2019 16:57, Waiman Long wrote:
>> It was found that passing an invalid cpu number to
pv_vcpu_is_preempted()
>> might panic the kernel in a VM guest. For example,
>>
>> [    2.531077] Oops: 0000 [#1] SMP PTI
>>   :
>> [    2.532545] Hardware name: Red Hat KVM, BIOS 0.5.1 01/01/2011
>> [    2.533321] RIP:
0010:__raw_callee_save___kvm_vcpu_is_preempted+0x0/0x20
>>
>> To guard against this kind of kernel panic, check is added to
>> pv_vcpu_is_preempted() to make sure that no invalid cpu number will
>> be used.
>>
>> Signed-off-by: Waiman Long <longman at redhat.com>
>> ---
>>  arch/x86/include/asm/paravirt.h | 6 ++++++
>>  1 file changed, 6 insertions(+)
>>
>> diff --git a/arch/x86/include/asm/paravirt.h
b/arch/x86/include/asm/paravirt.h
>> index c25c38a05c1c..4cfb465dcde4 100644
>> --- a/arch/x86/include/asm/paravirt.h
>> +++ b/arch/x86/include/asm/paravirt.h
>> @@ -671,6 +671,12 @@ static __always_inline void pv_kick(int cpu)
>>  
>>  static __always_inline bool pv_vcpu_is_preempted(long cpu)
>>  {
>> +	/*
>> +	 * Guard against invalid cpu number or the kernel might panic.
>> +	 */
>> +	if (WARN_ON_ONCE((unsigned long)cpu >= nr_cpu_ids))
>> +		return false;
>> +
>>  	return PVOP_CALLEE1(bool, lock.vcpu_is_preempted, cpu);
>>  }
> Can this really happen without being a programming error?
This shouldn't happen without a programming error, I think. In my case,
it was caused by a race condition leading to use-after-free of the cpu
number. However, my point is that error like that shouldn't cause the
kernel to panic.
> Basically you'd need to guard all percpu area accesses to foreign cpus
> this way. Why is this one special?
It depends. If out-of-bound access can only happen with obvious
programming error, I don't think we need to guard against them. In this
case, I am not totally sure if the race condition that I found may
happen with existing code or not. To be prudent, I decide to send this
patch out.

The race condition that I am looking at is as follows:

? CPU 0???????????????????????? CPU 1
? -----???????????????????????? -----
up_write:
? owner = NULL;
? <release-barrier>
? count = 0;

<rcu-free task structure>
?
????????????????????????? rwsem_can_spin_on_owner:
??????????????????????????? rcu_read_lock();
??????????????????????????? read owner;
????????????????????????????? :
??????????????????????????? vcpu_is_preempted(owner->cpu);
????????????????????????????? :
??????????????????????????? rcu_read_unlock()

When I tried to merge the owner into the count (clear the owner after
the barrier), I can reproduce the crash 100% when booting up the kernel
in a VM guest. However, I am not sure if the configuration above is safe
and is just very hard to reproduce.

Alternatively, I can also do the cpu check before calling
vcpu_is_preempted().

Cheers,
Longman

On 25/03/2019 19:03, Waiman Long wrote:
> On 03/25/2019 12:40 PM, Juergen Gross wrote:
>> On 25/03/2019 16:57, Waiman Long wrote:
>>> It was found that passing an invalid cpu number to
pv_vcpu_is_preempted()
>>> might panic the kernel in a VM guest. For example,
>>>
>>> [    2.531077] Oops: 0000 [#1] SMP PTI
>>>   :
>>> [    2.532545] Hardware name: Red Hat KVM, BIOS 0.5.1 01/01/2011
>>> [    2.533321] RIP:
0010:__raw_callee_save___kvm_vcpu_is_preempted+0x0/0x20
>>>
>>> To guard against this kind of kernel panic, check is added to
>>> pv_vcpu_is_preempted() to make sure that no invalid cpu number will
>>> be used.
>>>
>>> Signed-off-by: Waiman Long <longman at redhat.com>
>>> ---
>>>  arch/x86/include/asm/paravirt.h | 6 ++++++
>>>  1 file changed, 6 insertions(+)
>>>
>>> diff --git a/arch/x86/include/asm/paravirt.h
b/arch/x86/include/asm/paravirt.h
>>> index c25c38a05c1c..4cfb465dcde4 100644
>>> --- a/arch/x86/include/asm/paravirt.h
>>> +++ b/arch/x86/include/asm/paravirt.h
>>> @@ -671,6 +671,12 @@ static __always_inline void pv_kick(int cpu)
>>>  
>>>  static __always_inline bool pv_vcpu_is_preempted(long cpu)
>>>  {
>>> +	/*
>>> +	 * Guard against invalid cpu number or the kernel might panic.
>>> +	 */
>>> +	if (WARN_ON_ONCE((unsigned long)cpu >= nr_cpu_ids))
>>> +		return false;
>>> +
>>>  	return PVOP_CALLEE1(bool, lock.vcpu_is_preempted, cpu);
>>>  }
>> Can this really happen without being a programming error?
> 
> This shouldn't happen without a programming error, I think. In my case,
> it was caused by a race condition leading to use-after-free of the cpu
> number. However, my point is that error like that shouldn't cause the
> kernel to panic.
> 
>> Basically you'd need to guard all percpu area accesses to foreign
cpus
>> this way. Why is this one special?
> 
> It depends. If out-of-bound access can only happen with obvious
> programming error, I don't think we need to guard against them. In this
> case, I am not totally sure if the race condition that I found may
> happen with existing code or not. To be prudent, I decide to send this
> patch out.
> 
> The race condition that I am looking at is as follows:
> 
> ? CPU 0???????????????????????? CPU 1
> ? -----???????????????????????? -----
> up_write:
> ? owner = NULL;
> ? <release-barrier>
> ? count = 0;
> 
> <rcu-free task structure>
> ?
> ????????????????????????? rwsem_can_spin_on_owner:
> ??????????????????????????? rcu_read_lock();
> ??????????????????????????? read owner;
> ????????????????????????????? :
> ??????????????????????????? vcpu_is_preempted(owner->cpu);
> ????????????????????????????? :
> ??????????????????????????? rcu_read_unlock()
> 
> When I tried to merge the owner into the count (clear the owner after
> the barrier), I can reproduce the crash 100% when booting up the kernel
> in a VM guest. However, I am not sure if the configuration above is safe
> and is just very hard to reproduce.
> 
> Alternatively, I can also do the cpu check before calling
> vcpu_is_preempted().
I think I'd prefer that.


Juergen

On 04/01/2019 02:38 AM, Juergen Gross wrote:
> On 25/03/2019 19:03, Waiman Long wrote:
>> On 03/25/2019 12:40 PM, Juergen Gross wrote:
>>> On 25/03/2019 16:57, Waiman Long wrote:
>>>> It was found that passing an invalid cpu number to
pv_vcpu_is_preempted()
>>>> might panic the kernel in a VM guest. For example,
>>>>
>>>> [    2.531077] Oops: 0000 [#1] SMP PTI
>>>>   :
>>>> [    2.532545] Hardware name: Red Hat KVM, BIOS 0.5.1
01/01/2011
>>>> [    2.533321] RIP:
0010:__raw_callee_save___kvm_vcpu_is_preempted+0x0/0x20
>>>>
>>>> To guard against this kind of kernel panic, check is added to
>>>> pv_vcpu_is_preempted() to make sure that no invalid cpu number
will
>>>> be used.
>>>>
>>>> Signed-off-by: Waiman Long <longman at redhat.com>
>>>> ---
>>>>  arch/x86/include/asm/paravirt.h | 6 ++++++
>>>>  1 file changed, 6 insertions(+)
>>>>
>>>> diff --git a/arch/x86/include/asm/paravirt.h
b/arch/x86/include/asm/paravirt.h
>>>> index c25c38a05c1c..4cfb465dcde4 100644
>>>> --- a/arch/x86/include/asm/paravirt.h
>>>> +++ b/arch/x86/include/asm/paravirt.h
>>>> @@ -671,6 +671,12 @@ static __always_inline void pv_kick(int
cpu)
>>>>  
>>>>  static __always_inline bool pv_vcpu_is_preempted(long cpu)
>>>>  {
>>>> +	/*
>>>> +	 * Guard against invalid cpu number or the kernel might
panic.
>>>> +	 */
>>>> +	if (WARN_ON_ONCE((unsigned long)cpu >= nr_cpu_ids))
>>>> +		return false;
>>>> +
>>>>  	return PVOP_CALLEE1(bool, lock.vcpu_is_preempted, cpu);
>>>>  }
>>> Can this really happen without being a programming error?
>> This shouldn't happen without a programming error, I think. In my
case,
>> it was caused by a race condition leading to use-after-free of the cpu
>> number. However, my point is that error like that shouldn't cause
the
>> kernel to panic.
>>
>>> Basically you'd need to guard all percpu area accesses to
foreign cpus
>>> this way. Why is this one special?
>> It depends. If out-of-bound access can only happen with obvious
>> programming error, I don't think we need to guard against them. In
this
>> case, I am not totally sure if the race condition that I found may
>> happen with existing code or not. To be prudent, I decide to send this
>> patch out.
>>
>> The race condition that I am looking at is as follows:
>>
>> ? CPU 0???????????????????????? CPU 1
>> ? -----???????????????????????? -----
>> up_write:
>> ? owner = NULL;
>> ? <release-barrier>
>> ? count = 0;
>>
>> <rcu-free task structure>
>> ?
>> ????????????????????????? rwsem_can_spin_on_owner:
>> ??????????????????????????? rcu_read_lock();
>> ??????????????????????????? read owner;
>> ????????????????????????????? :
>> ??????????????????????????? vcpu_is_preempted(owner->cpu);
>> ????????????????????????????? :
>> ??????????????????????????? rcu_read_unlock()
>>
>> When I tried to merge the owner into the count (clear the owner after
>> the barrier), I can reproduce the crash 100% when booting up the kernel
>> in a VM guest. However, I am not sure if the configuration above is
safe
>> and is just very hard to reproduce.
>>
>> Alternatively, I can also do the cpu check before calling
>> vcpu_is_preempted().
> I think I'd prefer that.
>
>
> Juergen
>
It turns out that it may be caused by a software bug after all. You can
ignore this patch for now.

Thanks,
Longman