Re: [Qemu-devel] [PATCH] target-i386: Sanity check host processor physical address width

[Laszlo Ersek]

On 07/09/15 11:27, Igor Mammedov wrote:
> On Thu, 09 Jul 2015 09:02:38 +0200
> Laszlo Ersek <address@hidden> wrote:
> 
>> On 07/09/15 00:42, Bandan Das wrote:
>>>
>>> If a Linux guest is assigned more memory than is supported
>>> by the host processor, the guest is unable to boot. That
>>> is expected, however, there's no message indicating the user
>>> what went wrong. This change prints a message to stderr if
>>> KVM has the corresponding capability.
>>>
>>> Reported-by: Laszlo Ersek <address@hidden>
>>> Signed-off-by: Bandan Das <address@hidden>
>>> ---
>>>  linux-headers/linux/kvm.h | 1 +
>>>  target-i386/kvm.c         | 6 ++++++
>>>  2 files changed, 7 insertions(+)
>>>
>>> diff --git a/linux-headers/linux/kvm.h b/linux-headers/linux/kvm.h
>>> index 3bac873..6afad49 100644
>>> --- a/linux-headers/linux/kvm.h
>>> +++ b/linux-headers/linux/kvm.h
>>> @@ -817,6 +817,7 @@ struct kvm_ppc_smmu_info {
>>>  #define KVM_CAP_DISABLE_QUIRKS 116
>>>  #define KVM_CAP_X86_SMM 117
>>>  #define KVM_CAP_MULTI_ADDRESS_SPACE 118
>>> +#define KVM_CAP_PHY_ADDR_WIDTH 119
>>>  
>>>  #ifdef KVM_CAP_IRQ_ROUTING
>>>  
>>> diff --git a/target-i386/kvm.c b/target-i386/kvm.c
>>> index 066d03d..66e3448 100644
>>> --- a/target-i386/kvm.c
>>> +++ b/target-i386/kvm.c
>>> @@ -892,6 +892,7 @@ int kvm_arch_init(MachineState *ms, KVMState *s)
>>>      uint64_t shadow_mem;
>>>      int ret;
>>>      struct utsname utsname;
>>> +    int max_phys_bits;
>>>  
>>>      ret = kvm_get_supported_msrs(s);
>>>      if (ret < 0) {
>>> @@ -945,6 +946,11 @@ int kvm_arch_init(MachineState *ms, KVMState *s)
>>>          }
>>>      }
>>>  
>>> +    max_phys_bits = kvm_check_extension(s, KVM_CAP_PHY_ADDR_WIDTH);
>>> +    if (max_phys_bits && (1ULL << max_phys_bits) <= ram_size)
>>> +        fprintf(stderr, "Warning: The amount of memory assigned to the 
>>> guest "
>>> +            "is more than that supported by the host CPU(s). Guest may be 
>>> unstable.\n");
>>> +
>>>      if (kvm_check_extension(s, KVM_CAP_X86_SMM)) {
>>>          smram_machine_done.notify = register_smram_listener;
>>>          qemu_add_machine_init_done_notifier(&smram_machine_done);
>>>
>>
>> First, see my comments on the KVM patch.
>>
>> Second, ram_size is not the right thing to compare. What should be
>> checked is whether the highest guest-physical address that maps to RAM
>> can be represented in the address width of the host processor (and only
>> if EPT is enabled, but that sub-condition belongs to the KVM patch).
>>
>> Note that this is not the same as the check written in the patch. For
>> example, if you assume a 32-bit PCI hole with size 1 GB, then a total
>> guest RAM of size 63 GB will result in the highest guest-phys memory
>> address being 0xF_FFFF_FFFF, which just fits into 36 bits.
>>
>> Correspondingly, the above code would not print the warning for
>>
>>   -m $((63 * 1024 + 1))
>>
>> on my laptop (which has "address sizes   : 36 bits physical, ..."), even
>> though such a guest would not boot for me (with EPT enabled).
>>
>> Please see
>>
>> http://thread.gmane.org/gmane.comp.bios.tianocore.devel/15418/focus=15447
>>
>> So, "ram_size" in the controlling expression should be replaced with
>> "maximum_guest_ram_address" (which should be inclusive, and the <= relop
>> should be preserved).
> also with memory hotplug tuned on we should check if the end of
> hotplug memory area is less then limit, i.e.:
> 
>   pcms->hotplug_memory.base + hotplug_mem_size < 1ULL << max_phys_bits

Seems reasonable, thanks for the hint!

(The LHS in this instance is exclusive though, so equality should *not*
trigger the warning. "maximum_guest_ram_address" is inclusive, and
equality should trigger the warning. (Although equality seems quite
impossible in practice.))

Thanks!
Laszlo