Re: [patch] kvm: suppress KVM_SET_GSI_ROUTING allocation failure

[Christian Borntraeger]

On 02/14/2018 02:03 AM, David Rientjes wrote:
> On Tue, 13 Feb 2018, Paolo Bonzini wrote:
> 
>>>> The KVM_SET_GSI_ROUTING ioctl does a vmalloc() of
>>>> sizeof(struct kvm_irq_routing_entry) multiplied by a user-supplied value.
>>>> This can be up to 4096 entries on architectures such as arm64 and s390
>>>> (and the upper bound may be increased on s390 eventually).
>>>>
>>>> This can produce a vmalloc allocation failure warning:
>>>>
>>> [...]
>>>>  kvm_vm_ioctl+0x910/0x15e0 arch/x86/kvm/../../../virt/kvm/kvm_main.c:4153
>>>
>>>                                                                        ^^^^^
>>>
>>>> @@ -3063,7 +3063,8 @@ static long kvm_vm_ioctl(struct file *filp,
>>>
>>>       ^^^^^
>>>
>>>
>>> Are you sure that you got the right vmalloc?
>>
>> Nice catch!  But well, it's the only one in the whole file. :)
>>
>> That seems very much like an old patch then.  I'm unqueuing it.
>>
> 
> It's not a catch at all, the fact that I saw this warning with an older 
> kernel for KVM_SET_GSI_ROUTING doesn't mean that I can't patch it with an 
> upstream kernel.  Would you prefer I remove the stack trace completely?

FWIW, your stack trace did not complain about a too big allocation, it
complained about 0 allocation:

----- snip ------
vmalloc: allocation failure: 0 bytes, mode:0x24000c2(GFP_KERNEL|__GFP_HIGHMEM)
----- snip ------

After commit f8c1b85b2523 ("KVM: x86: avoid vmalloc(0) in the KVM_SET_CPUID)"
this case should be prevented. The only question is does your patch makes sense
nevertheless as we gracefully handle the ENOMEM case? So a reproducer on
a newer kernel would be good. Maybe use the "vmalloc" kernel parameter to force
this.