On 18.11.21 09:47, Jan Beulich wrote:
> On 18.11.2021 06:32, Juergen Gross wrote:
>> On 18.11.21 03:37, Stefano Stabellini wrote:
>>> --- a/drivers/xen/xenbus/xenbus_probe.c
>>> +++ b/drivers/xen/xenbus/xenbus_probe.c
>>> @@ -951,6 +951,28 @@ static int __init xenbus_init(void)
>>>    		err = hvm_get_parameter(HVM_PARAM_STORE_PFN, &v);
>>>    		if (err)
>>>    			goto out_error;
>>> +		/*
>>> +		 * Return error on an invalid value.
>>> +		 *
>>> +		 * Uninitialized hvm_params are zero and return no error.
>>> +		 * Although it is theoretically possible to have
>>> +		 * HVM_PARAM_STORE_PFN set to zero on purpose, in reality it is
>>> +		 * not zero when valid. If zero, it means that Xenstore hasn't
>>> +		 * been properly initialized. Instead of attempting to map a
>>> +		 * wrong guest physical address return error.
>>> +		 */
>>> +		if (v == 0) {
>>
>> Make this "if (v == ULONG_MAX || v== 0)" instead?
>> This would result in the same err on a new and an old hypervisor
>> (assuming we switch the hypervisor to init params with ~0UL).
>>
>>> +			err = -ENOENT;
>>> +			goto out_error;
>>> +		}
>>> +		/*
>>> +		 * ULONG_MAX is invalid on 64-bit because is INVALID_PFN.
>>> +		 * On 32-bit return error to avoid truncation.
>>> +		 */
>>> +		if (v >= ULONG_MAX) {
>>> +			err = -EINVAL;
>>> +			goto out_error;
>>> +		}
>>
>> Does it make sense to continue the system running in case of
>> truncation? This would be a 32-bit guest with more than 16TB of RAM
>> and the Xen tools decided to place the Xenstore ring page above the
>> 16TB boundary. This is a completely insane scenario IMO.
>>
>> A proper panic() in this case would make diagnosis of that much
>> easier (me having doubts that this will ever be hit, though).
> 
> While I agree panic() may be an option here (albeit I'm not sure why
> that would be better than trying to cope with 0 and hence without

I could imagine someone wanting to run a guest without Xenstore access,
which BTW will happen in case of a guest created by the hypervisor at
boot time.

> xenbus), I'd like to point out that the amount of RAM assigned to a
> guest is unrelated to the choice of GFNs for the various "magic"
> items.

Yes, but this would still be a major tools problem which probably
would render the whole guest rather unusable.


Juergen