Re: [PATCH 1/2] x86: Make hvm_enabled a constant unless both PV and HVM are both compiled

[Alejandro Vallejo <alejandro.garciavallejo@amd.com>]

On Fri Feb 13, 2026 at 3:26 PM CET, Roger Pau Monné wrote:
> On Fri, Feb 13, 2026 at 02:37:29PM +0100, Alejandro Vallejo wrote:
>> PV-shim already has hvm_enabled optimised to false, but there's no
>> reason HVM-only builds can't benefit from the same optimisation as long
>> as we add a boot-time panic in case HVM support is missed. Many branches
>> go away, some in the potentially hot switch_cr3_cr4.
>> 
>> HVM-only:
>>   add/remove: 0/1 grow/shrink: 1/9 up/down: 1/-162 (-161)
>>   Function                                     old     new   delta
>>   arch_do_physinfo                              79      80      +1
>>   hvm_enabled                                    1       -      -1
>>   symbols_offsets                            30732   30728      -4
>>   symbols_names                             108029  108022      -7
>>   symbols_sorted_offsets                     60656   60648      -8
>>   flush_area_local                             571     562      -9
>>   switch_cr3_cr4                               311     300     -11
>>   init_xen_cap_info                             62      43     -19
>>   arch_sanitise_domain_config                  885     863     -22
>>   init_guest_cpu_policies                     1270    1247     -23
>>   hvm_domain_initialise                       1127    1069     -58
>>   Total: Before=3797004, After=3796843, chg -0.00%
>> 
>> With hvm_enabled const-ified, it's fine to take hvm_flush_guest_tlbs()
>> outside the CONFIG_HVM ifdef and remove the stub. They compile to the
>> same code after DCE.
>> 
>> Signed-off-by: Alejandro Vallejo <alejandro.garciavallejo@amd.com>
>> ---
>>  xen/arch/x86/hvm/hvm.c             |  9 +++++++++
>>  xen/arch/x86/include/asm/hvm/hvm.h | 30 +++++++++++++++---------------
>>  2 files changed, 24 insertions(+), 15 deletions(-)
>> 
>> diff --git a/xen/arch/x86/hvm/hvm.c b/xen/arch/x86/hvm/hvm.c
>> index 4d37a93c57..da56944e74 100644
>> --- a/xen/arch/x86/hvm/hvm.c
>> +++ b/xen/arch/x86/hvm/hvm.c
>> @@ -72,7 +72,9 @@
>>  
>>  #include <compat/hvm/hvm_op.h>
>>  
>> +#ifdef CONFIG_PV
>>  bool __read_mostly hvm_enabled;
>
> __ro_after_init?

Yeah, seems to have been missing originally

>
>> +#endif /* CONFIG_PV */
>>  
>>  #ifdef DBG_LEVEL_0
>>  unsigned int opt_hvm_debug_level __read_mostly;
>> @@ -173,9 +175,16 @@ static int __init cf_check hvm_enable(void)
>>          svm_fill_funcs();
>>  
>>      if ( fns == NULL )
>> +    {
>> +        if ( !IS_ENABLED(CONFIG_PV) )
>> +            panic("HVM support not detected and PV compiled-out\n");
>> +
>>          return 0;
>> +    }
>>  
>> +#ifdef CONFIG_PV
>
> CONFIG_HVM I think?

No. CONFIG_HVM always holds here, what we want to remove is hvm_enabled being
present when CONFIG_PV is _also_ present.

>
>>      hvm_enabled = 1;
>
> = true;

True enough.

>
>> +#endif /* CONFIG_PV */
>>  
>>      printk("HVM: %s enabled\n", fns->name);
>>      if ( !hap_supported(&hvm_funcs) )
>> diff --git a/xen/arch/x86/include/asm/hvm/hvm.h b/xen/arch/x86/include/asm/hvm/hvm.h
>> index 7d9774df59..dc609bf4cb 100644
>> --- a/xen/arch/x86/include/asm/hvm/hvm.h
>> +++ b/xen/arch/x86/include/asm/hvm/hvm.h
>> @@ -261,7 +261,11 @@ struct hvm_function_table {
>>  };
>>  
>>  extern struct hvm_function_table hvm_funcs;
>> +#if defined(CONFIG_PV) && defined(CONFIG_HVM)
>>  extern bool hvm_enabled;
>> +#else
>> +#define hvm_enabled IS_ENABLED(CONFIG_HVM)
>> +#endif
>>  extern int8_t hvm_port80_allowed;
>>  
>>  extern const struct hvm_function_table *start_svm(void);
>> @@ -399,6 +403,17 @@ static inline bool using_svm(void)
>>  #define hvm_is_in_uc_mode(d) \
>>      (using_vmx() && (d)->arch.hvm.vmx.in_uc_mode)
>>  
>> +/*
>> + * Called to ensure than all guest-specific mappings in a tagged TLB are
>> + * flushed; does *not* flush Xen's TLB entries, and on processors without a
>> + * tagged TLB it will be a noop.
>> + */
>> +static inline void hvm_flush_guest_tlbs(void)
>> +{
>> +    if ( hvm_enabled )
>> +        hvm_asid_flush_core();
>> +}
>> +
>>  #ifdef CONFIG_HVM
>>  
>>  #define hvm_get_guest_tsc(v) hvm_get_guest_tsc_fixed(v, 0)
>> @@ -498,17 +513,6 @@ static inline void hvm_set_tsc_offset(struct vcpu *v, uint64_t offset)
>>      alternative_vcall(hvm_funcs.set_tsc_offset, v, offset);
>>  }
>>  
>> -/*
>> - * Called to ensure than all guest-specific mappings in a tagged TLB are 
>> - * flushed; does *not* flush Xen's TLB entries, and on processors without a 
>> - * tagged TLB it will be a noop.
>> - */
>> -static inline void hvm_flush_guest_tlbs(void)
>> -{
>> -    if ( hvm_enabled )
>> -        hvm_asid_flush_core();
>> -}
>
> Is there any specific reason you only pick hvm_flush_guest_tlbs().

I didn't try to remove more. That one was the only one with hvm_enabled in the
static inline so it seems easy to pick apart.

I just tried compiling and I require very few additions to make the build
compile without stubs. I'll send another version with the adjustments needed.

Cheers,
Alejandro