Re: [PATCH 1/2] x86: correct is_pv_domain() when !CONFIG_PV

[Jan Beulich <jbeulich@suse.com>]

On 12.04.2021 16:49, Roger Pau Monné wrote:
> On Mon, Apr 12, 2021 at 12:07:12PM +0200, Jan Beulich wrote:
>> On 12.04.2021 11:34, Roger Pau Monné wrote:
>>> On Fri, Nov 27, 2020 at 05:54:57PM +0100, Jan Beulich wrote:
>>>> --- a/xen/arch/x86/dom0_build.c
>>>> +++ b/xen/arch/x86/dom0_build.c
>>>> @@ -568,7 +568,7 @@ int __init construct_dom0(struct domain
>>>>  
>>>>      if ( is_hvm_domain(d) )
>>>>          rc = dom0_construct_pvh(d, image, image_headroom, initrd, cmdline);
>>>> -    else if ( is_pv_domain(d) )
>>>> +    else if ( is_pv_64bit_domain(d) || is_pv_32bit_domain(d) )
>>>
>>> Urg, that's very confusing IMO, as I'm sure I would ask someone to
>>> just use is_pv_domain without realizing. It needs at least a comment,
>>> but even then I'm not sure I like it.
>>
>> I can add a comment, sure, but I think this is as confusing (or not)
>> as ...
>>
>>> So that I understand it, the point to use those expressions instead of
>>> is_pv_domain is to avoid calling dom0_construct_pv when CONFIG_PV is
>>> not enabled?
>>>
>>> Maybe it wold be better to instead use:
>>>
>>> if ( IS_ENABLED(CONFIG_PV) && is_pv_domain(d) )
>>
>> ... this.
>>
>>> In any case I wonder if we should maybe aim to introduce a new type
>>> for system domains, that's neither PV or HVM, in order to avoid having
>>> system domains qualified as PV even when PV is compiled out.
>>
>> This was my first thought, too, but would come with a much higher
>> price tag: We'd need to audit all uses for whether they're meant
>> to include the special domains. And this includes auditing of cases
>> where !is_hvm_*() may be inferred to mean is_pv_*().
> 
> What about we provide a dummy dom0_construct_pv that returns
> -EOPNOTSUPP when !CONFIG_PV and take rc into account for the panic
> call in construct_dom0 ie:
> 
>     if ( is_hvm_domain(d) )
>         rc = dom0_construct_pvh(d, image, image_headroom, initrd, cmdline);
>     else
>         rc = dom0_construct_pv(d, image, image_headroom, initrd, cmdline);
> 
>     if ( rc == -EOPNOTSUPP )
>         panic("Cannot construct Dom0. No guest interface available\n");
>     if ( rc )
>         return rc;
> 
> I think that's likely less confusing that the alternatives.

This could certainly be made work, but see below (i.e. it would
help the situation right here, but not the general issue - the
case in arch_do_domctl() may look less confusing, but really
suffers the same problem).

>>>> --- a/xen/arch/x86/domain.c
>>>> +++ b/xen/arch/x86/domain.c
>>>> @@ -1544,6 +1544,7 @@ arch_do_vcpu_op(
>>>>   */
>>>>  static void load_segments(struct vcpu *n)
>>>>  {
>>>> +#ifdef CONFIG_PV
>>>>      struct cpu_user_regs *uregs = &n->arch.user_regs;
>>>>      unsigned long gsb = 0, gss = 0;
>>>>      bool compat = is_pv_32bit_vcpu(n);
>>>> @@ -1709,6 +1710,7 @@ static void load_segments(struct vcpu *n
>>>>          regs->cs            = FLAT_KERNEL_CS;
>>>>          regs->rip           = pv->failsafe_callback_eip;
>>>>      }
>>>> +#endif
>>>>  }
>>>>  
>>>>  /*
>>>> @@ -1723,6 +1725,7 @@ static void load_segments(struct vcpu *n
>>>>   */
>>>>  static void save_segments(struct vcpu *v)
>>>>  {
>>>> +#ifdef CONFIG_PV
>>>>      struct cpu_user_regs *regs = &v->arch.user_regs;
>>>>  
>>>>      read_sregs(regs);
>>>> @@ -1748,6 +1751,7 @@ static void save_segments(struct vcpu *v
>>>>          else
>>>>              v->arch.pv.gs_base_user = gs_base;
>>>>      }
>>>> +#endif
>>>>  }
>>>
>>> Could you move {load,save}_segments to pv/domain.c and rename to
>>> pv_{load,save}_segments and provide a dummy handler for !CONFIG_PV in
>>> pv/domain.h?
>>>
>>> Sorry it's slightly more work, but I think it's cleaner overall.
>>
>> Doing so was my first thought too, but we'd lose the present inlining
>> of the functions. For save_segments() this could be dealt with by
>> moving paravirt_ctxt_switch_from() as well, but load_segments() would
>> remain.
> 
> I see, maybe worth marking as inline then or adding a note about why
> they are not moved to pv/domain.c?

We try to avoid marking functions inline outside of headers. Adding
a note is an option, but I'm not sure something to be done here.

> As an aside, why do we need to call load_segments with interrupts
> enabled? Could we move it to paravirt_ctxt_switch_to?

load_segments() can raise faults, and faults with interrupts
disabled are, with (intentionally) very few exceptions, fatal.

>>>> --- a/xen/include/xen/sched.h
>>>> +++ b/xen/include/xen/sched.h
>>>> @@ -985,7 +985,7 @@ static always_inline bool is_control_dom
>>>>  
>>>>  static always_inline bool is_pv_domain(const struct domain *d)
>>>>  {
>>>> -    return IS_ENABLED(CONFIG_PV) &&
>>>> +    return IS_ENABLED(CONFIG_X86) &&
>>>>          evaluate_nospec(!(d->options & XEN_DOMCTL_CDF_hvm));
>>>>  }
>>>>  
>>>> @@ -1011,7 +1011,7 @@ static always_inline bool is_pv_32bit_vc
>>>>  
>>>>  static always_inline bool is_pv_64bit_domain(const struct domain *d)
>>>>  {
>>>> -    if ( !is_pv_domain(d) )
>>>> +    if ( !IS_ENABLED(CONFIG_PV) || !is_pv_domain(d) )
>>>>          return false;
>>>
>>> I think overall is confusing to have a domain that returns true for
>>> is_pv_domain but false for both is_pv_{64,32}bit_domain checks.
>>>
>>> I know those are only the system domains, but it feels confusing and
>>> could cause mistakes in the future IMO, as then we would have to
>>> carefully think where to use ( is_pv_64bit_domain(d)
>>> || is_pv_32bit_domain(d) ) vs just using is_pv_domain(d), or
>>> IS_ENABLED(CONFIG_PV) && is_pv_domain(d)
>>
>> Imo it's not "then we would have to carefully think where to use ..."
>> but instead this patch is an indication that we should have been for
>> quite some time. For this reason (coming back to your first comment
>> at the top) I'm not sure adding a comment _there_ is actually useful.
>> Every use of is_pv_*() needs carefully considering which domains are
>> really meant.
> 
> Maybe we shouldn't have used is_pv_domain as a way to hide code from
> the compiler and instead always provide dummy functions, as even with
> PV support compiled out we still need some of it for system domains.
> 
> I'm not sure I have a good proposal to make, but it seems wrong to me
> that is_pv_domain(d) could be different than is_pv_64bit_domain(d) ||
> is_pv_32bit_domain(d).

Hmm, so we're of opposite opinions - not sure what to do. Short of
having / introducing is_system_domain() or some such (with all the
needed auditing) I can't see how assuming the two would mean the
same could ever have been true. With what we have is_pv_domain() is
legitimately true for them, and both is_pv_{32,64}bit_domain() ought
to be false (as there's no specific bitness associated with them)
imo _at least_ when !PV.

Jan