x86: Clear LAM and FRED feature bits

[PATCH] x86: Clear LAM and FRED feature bits

Posted by Maciej Wieczor-Retman 2 months, 2 weeks ago

If LAM (CONFIG_ADDRESS_MASKING) or FRED (CONFIG_X86_FRED) are disabled
during compile time, they still are enumerated in macros such as
cpu_has() or this_cpu_has() that use the x86_capability bitmask. The
features are also visible in /proc/cpuinfo even though they are not
enabled - which is contrary to what the documentation states about the
file.

The documentation for x86/cpuinfo.rst also states how to handle features
disabled at compile time:

	The kernel disabled support for it at compile-time
	--------------------------------------------------

	For example, if Linear Address Masking (LAM) is not enabled when
	building (i.e., CONFIG_ADDRESS_MASKING is not selected) the flag
	"lam" will not show up. Even though the feature will still be
	detected via CPUID, the kernel disables it by clearing via
	setup_clear_cpu_cap(X86_FEATURE_LAM).

Clear feature bits if they are present in the DISABLED_MASK.

Signed-off-by: Maciej Wieczor-Retman <maciej.wieczor-retman@intel.com>
---
 arch/x86/kernel/cpu/common.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
index 77afca95cced..1c5af795cedc 100644
--- a/arch/x86/kernel/cpu/common.c
+++ b/arch/x86/kernel/cpu/common.c
@@ -1782,6 +1782,16 @@ static void __init early_identify_cpu(struct cpuinfo_x86 *c)
 	if (!pgtable_l5_enabled())
 		setup_clear_cpu_cap(X86_FEATURE_LA57);
 
+	/*
+	 * If a feature is disabled during compile time clear its feature
+	 * bit to prevent it from showing up in the x86_capability bitmask.
+	 */
+	if (!cpu_feature_enabled(X86_FEATURE_LAM))
+		setup_clear_cpu_cap(X86_FEATURE_LAM);
+
+	if (!cpu_feature_enabled(X86_FEATURE_FRED))
+		setup_clear_cpu_cap(X86_FEATURE_FRED);
+
 	detect_nopl();
 }
 
-- 
2.49.0

Re: [PATCH] x86: Clear LAM and FRED feature bits

Posted by Xin Li 2 months, 2 weeks ago

> On Jul 22, 2025, at 12:54 AM, Maciej Wieczor-Retman <maciej.wieczor-retman@intel.com> wrote:
> 
> If LAM (CONFIG_ADDRESS_MASKING) or FRED (CONFIG_X86_FRED) are disabled
> during compile time, they still are enumerated in macros such as
> cpu_has() or this_cpu_has() that use the x86_capability bitmask. The
> features are also visible in /proc/cpuinfo even though they are not
> enabled - which is contrary to what the documentation states about the
> file.

I'm curious how this bug is introduced with my patch set that adds the
AWK script which automatically generates the CPU feature mask header.

> 
> The documentation for x86/cpuinfo.rst also states how to handle features
> disabled at compile time:
> 
>    The kernel disabled support for it at compile-time
>    --------------------------------------------------
> 
>    For example, if Linear Address Masking (LAM) is not enabled when
>    building (i.e., CONFIG_ADDRESS_MASKING is not selected) the flag
>    "lam" will not show up. Even though the feature will still be
>    detected via CPUID, the kernel disables it by clearing via
>    setup_clear_cpu_cap(X86_FEATURE_LAM).
> 
> Clear feature bits if they are present in the DISABLED_MASK.
> 
> Signed-off-by: Maciej Wieczor-Retman <maciej.wieczor-retman@intel.com>

And we need a Fixes tag and CC stable.

> ---
> arch/x86/kernel/cpu/common.c | 10 ++++++++++
> 1 file changed, 10 insertions(+)
> 
> diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
> index 77afca95cced..1c5af795cedc 100644
> --- a/arch/x86/kernel/cpu/common.c
> +++ b/arch/x86/kernel/cpu/common.c
> @@ -1782,6 +1782,16 @@ static void __init early_identify_cpu(struct cpuinfo_x86 *c)
>    if (!pgtable_l5_enabled())
>        setup_clear_cpu_cap(X86_FEATURE_LA57);
> 
> +    /*
> +     * If a feature is disabled during compile time clear its feature
> +     * bit to prevent it from showing up in the x86_capability bitmask.
> +     */
> +    if (!cpu_feature_enabled(X86_FEATURE_LAM))
> +        setup_clear_cpu_cap(X86_FEATURE_LAM);
> +
> +    if (!cpu_feature_enabled(X86_FEATURE_FRED))
> +        setup_clear_cpu_cap(X86_FEATURE_FRED);
> +

The following code will work as a generic fix:

	c->x86_capability[i] &= ~DISABLED_MASK(i);

And DISABLED_MASK(x) needs to be defined like DISABLED_MASK_BIT_SET(x).

Thanks!
     Xin

Re: [PATCH] x86: Clear LAM and FRED feature bits

Posted by H. Peter Anvin 2 months, 2 weeks ago

On July 22, 2025 2:08:28 AM PDT, Xin Li <xin@zytor.com> wrote:
>
>> On Jul 22, 2025, at 12:54 AM, Maciej Wieczor-Retman <maciej.wieczor-retman@intel.com> wrote:
>> 
>> If LAM (CONFIG_ADDRESS_MASKING) or FRED (CONFIG_X86_FRED) are disabled
>> during compile time, they still are enumerated in macros such as
>> cpu_has() or this_cpu_has() that use the x86_capability bitmask. The
>> features are also visible in /proc/cpuinfo even though they are not
>> enabled - which is contrary to what the documentation states about the
>> file.
>
>I'm curious how this bug is introduced with my patch set that adds the
>AWK script which automatically generates the CPU feature mask header.
>
>> 
>> The documentation for x86/cpuinfo.rst also states how to handle features
>> disabled at compile time:
>> 
>>    The kernel disabled support for it at compile-time
>>    --------------------------------------------------
>> 
>>    For example, if Linear Address Masking (LAM) is not enabled when
>>    building (i.e., CONFIG_ADDRESS_MASKING is not selected) the flag
>>    "lam" will not show up. Even though the feature will still be
>>    detected via CPUID, the kernel disables it by clearing via
>>    setup_clear_cpu_cap(X86_FEATURE_LAM).
>> 
>> Clear feature bits if they are present in the DISABLED_MASK.
>> 
>> Signed-off-by: Maciej Wieczor-Retman <maciej.wieczor-retman@intel.com>
>
>And we need a Fixes tag and CC stable.
>
>> ---
>> arch/x86/kernel/cpu/common.c | 10 ++++++++++
>> 1 file changed, 10 insertions(+)
>> 
>> diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
>> index 77afca95cced..1c5af795cedc 100644
>> --- a/arch/x86/kernel/cpu/common.c
>> +++ b/arch/x86/kernel/cpu/common.c
>> @@ -1782,6 +1782,16 @@ static void __init early_identify_cpu(struct cpuinfo_x86 *c)
>>    if (!pgtable_l5_enabled())
>>        setup_clear_cpu_cap(X86_FEATURE_LA57);
>> 
>> +    /*
>> +     * If a feature is disabled during compile time clear its feature
>> +     * bit to prevent it from showing up in the x86_capability bitmask.
>> +     */
>> +    if (!cpu_feature_enabled(X86_FEATURE_LAM))
>> +        setup_clear_cpu_cap(X86_FEATURE_LAM);
>> +
>> +    if (!cpu_feature_enabled(X86_FEATURE_FRED))
>> +        setup_clear_cpu_cap(X86_FEATURE_FRED);
>> +
>
>The following code will work as a generic fix:
>
>	c->x86_capability[i] &= ~DISABLED_MASK(i);
>
>And DISABLED_MASK(x) needs to be defined like DISABLED_MASK_BIT_SET(x).
>
>Thanks!
>    Xin

The easiest thing would be to initialize the setup disabled mask with the DISABLED bitmask instead of zero. This can be done statically; if it isn't already the awk script can produce the disabled bitmask in array form.

Re: [PATCH] x86: Clear LAM and FRED feature bits

Posted by Maciej Wieczor-Retman 2 months, 2 weeks ago

On 2025-07-22 at 09:46:42 -0700, H. Peter Anvin wrote:
>On July 22, 2025 2:08:28 AM PDT, Xin Li <xin@zytor.com> wrote:
>>> On Jul 22, 2025, at 12:54 AM, Maciej Wieczor-Retman <maciej.wieczor-retman@intel.com> wrote:
>>> ---
>>> arch/x86/kernel/cpu/common.c | 10 ++++++++++
>>> 1 file changed, 10 insertions(+)
>>> 
>>> diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
>>> index 77afca95cced..1c5af795cedc 100644
>>> --- a/arch/x86/kernel/cpu/common.c
>>> +++ b/arch/x86/kernel/cpu/common.c
>>> @@ -1782,6 +1782,16 @@ static void __init early_identify_cpu(struct cpuinfo_x86 *c)
>>>    if (!pgtable_l5_enabled())
>>>        setup_clear_cpu_cap(X86_FEATURE_LA57);
>>> 
>>> +    /*
>>> +     * If a feature is disabled during compile time clear its feature
>>> +     * bit to prevent it from showing up in the x86_capability bitmask.
>>> +     */
>>> +    if (!cpu_feature_enabled(X86_FEATURE_LAM))
>>> +        setup_clear_cpu_cap(X86_FEATURE_LAM);
>>> +
>>> +    if (!cpu_feature_enabled(X86_FEATURE_FRED))
>>> +        setup_clear_cpu_cap(X86_FEATURE_FRED);
>>> +
>>
>>The following code will work as a generic fix:
>>
>>	c->x86_capability[i] &= ~DISABLED_MASK(i);
>>
>>And DISABLED_MASK(x) needs to be defined like DISABLED_MASK_BIT_SET(x).
>>
>>Thanks!
>>    Xin
>
>The easiest thing would be to initialize the setup disabled mask with the
>DISABLED bitmask instead of zero. This can be done statically; if it isn't
>already the awk script can produce the disabled bitmask in array form.

Oh, that does sound simple. I'll try that too, see what ends up with cleaner
code.

Thanks :)

-- 
Kind regards
Maciej Wieczór-Retman

Re: [PATCH] x86: Clear LAM and FRED feature bits

Posted by H. Peter Anvin 2 months, 2 weeks ago

On 2025-07-22 11:19, Maciej Wieczor-Retman wrote:
>>
>> The easiest thing would be to initialize the setup disabled mask with the
>> DISABLED bitmask instead of zero. This can be done statically; if it isn't
>> already the awk script can produce the disabled bitmask in array form.
> 
> Oh, that does sound simple. I'll try that too, see what ends up with cleaner
> code.
> 

Xin should have code from me for exactly this.

	-hpa

Re: [PATCH] x86: Clear LAM and FRED feature bits

Posted by Xin Li 2 months, 2 weeks ago

On 7/22/2025 9:46 AM, H. Peter Anvin wrote:
>> The following code will work as a generic fix:
>>
>> 	c->x86_capability[i] &= ~DISABLED_MASK(i);
>>
>> And DISABLED_MASK(x) needs to be defined like DISABLED_MASK_BIT_SET(x).
>>
>> Thanks!
>>     Xin
> The easiest thing would be to initialize the setup disabled mask with the DISABLED bitmask instead of zero. This can be done statically; if it isn't already the awk script can produce the disabled bitmask in array form.

Yes, something like:

	void __init init_cpu_cap(void)
	{
		for (i = 0; i < NCAPINTS; i++) {
			cpu_caps_set[i] = REQUIRED_MASK(i);
			cpu_caps_cleared[i] = DISABLED_MASK(i);
		}
	}

And it would be better if it could be done at build time (to avoid
changing Xen which has a dedicated startup code path):

	__u32 cpu_caps_{set,cleared}[NCAPINTS + NBUGINTS] = {
		{REQUIRED,DISABLED}_MASK(i),
	};

And then apply_forced_caps() will do the rest automatically :)

Re: [PATCH] x86: Clear LAM and FRED feature bits

Posted by Maciej Wieczor-Retman 2 months, 2 weeks ago

On 2025-07-22 at 02:08:28 -0700, Xin Li wrote:
>
>> On Jul 22, 2025, at 12:54 AM, Maciej Wieczor-Retman <maciej.wieczor-retman@intel.com> wrote:
>> 
>> If LAM (CONFIG_ADDRESS_MASKING) or FRED (CONFIG_X86_FRED) are disabled
>> during compile time, they still are enumerated in macros such as
>> cpu_has() or this_cpu_has() that use the x86_capability bitmask. The
>> features are also visible in /proc/cpuinfo even though they are not
>> enabled - which is contrary to what the documentation states about the
>> file.
>
>I'm curious how this bug is introduced with my patch set that adds the
>AWK script which automatically generates the CPU feature mask header.

While your patch works great in setting up the DISABLED_MASKs, there is no place
that would read that mask and clear x86_capability bitmask. Most other features
seem to have a call to setup_clear_cpu_cap() somewhere, or they're in the
required bitmask. LAM doesn't have one at all, and FRED only has a kernel
command line check - so no compile-time support is checked anywhere.

>
>> 
>> The documentation for x86/cpuinfo.rst also states how to handle features
>> disabled at compile time:
>> 
>>    The kernel disabled support for it at compile-time
>>    --------------------------------------------------
>> 
>>    For example, if Linear Address Masking (LAM) is not enabled when
>>    building (i.e., CONFIG_ADDRESS_MASKING is not selected) the flag
>>    "lam" will not show up. Even though the feature will still be
>>    detected via CPUID, the kernel disables it by clearing via
>>    setup_clear_cpu_cap(X86_FEATURE_LAM).
>> 
>> Clear feature bits if they are present in the DISABLED_MASK.
>> 
>> Signed-off-by: Maciej Wieczor-Retman <maciej.wieczor-retman@intel.com>
>
>And we need a Fixes tag and CC stable.

Ah, thank you for reminding me :)

>
>> ---
>> arch/x86/kernel/cpu/common.c | 10 ++++++++++
>> 1 file changed, 10 insertions(+)
>> 
>> diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
>> index 77afca95cced..1c5af795cedc 100644
>> --- a/arch/x86/kernel/cpu/common.c
>> +++ b/arch/x86/kernel/cpu/common.c
>> @@ -1782,6 +1782,16 @@ static void __init early_identify_cpu(struct cpuinfo_x86 *c)
>>    if (!pgtable_l5_enabled())
>>        setup_clear_cpu_cap(X86_FEATURE_LA57);
>> 
>> +    /*
>> +     * If a feature is disabled during compile time clear its feature
>> +     * bit to prevent it from showing up in the x86_capability bitmask.
>> +     */
>> +    if (!cpu_feature_enabled(X86_FEATURE_LAM))
>> +        setup_clear_cpu_cap(X86_FEATURE_LAM);
>> +
>> +    if (!cpu_feature_enabled(X86_FEATURE_FRED))
>> +        setup_clear_cpu_cap(X86_FEATURE_FRED);
>> +
>
>The following code will work as a generic fix:
>
>	c->x86_capability[i] &= ~DISABLED_MASK(i);
>
>And DISABLED_MASK(x) needs to be defined like DISABLED_MASK_BIT_SET(x).

Thanks, I'll add a helper for that and put the call in the same place this patch
changes are. Seems like a good place with a lot of other setup_clear_cpu_cap()
calls.

>
>Thanks!
>    Xin

-- 
Kind regards
Maciej Wieczór-Retman

Re: [PATCH] x86: Clear LAM and FRED feature bits

Posted by Dave Hansen 2 months, 2 weeks ago

On 7/22/25 02:08, Xin Li wrote:
>> +    /*
>> +     * If a feature is disabled during compile time clear its feature
>> +     * bit to prevent it from showing up in the x86_capability bitmask.
>> +     */
>> +    if (!cpu_feature_enabled(X86_FEATURE_LAM))
>> +        setup_clear_cpu_cap(X86_FEATURE_LAM);
>> +
>> +    if (!cpu_feature_enabled(X86_FEATURE_FRED))
>> +        setup_clear_cpu_cap(X86_FEATURE_FRED);
>> +
> 
> The following code will work as a generic fix:
> 
>     c->x86_capability[i] &= ~DISABLED_MASK(i);
> 
> And DISABLED_MASK(x) needs to be defined like DISABLED_MASK_BIT_SET(x).

Maciej, I would much rather have a generic fix than force everyone to
remember to open-code this for every new feature that gets disabled.

Re: [PATCH] x86: Clear LAM and FRED feature bits

Posted by Maciej Wieczor-Retman 2 months, 2 weeks ago

On 2025-07-22 at 07:11:47 -0700, Dave Hansen wrote:
>On 7/22/25 02:08, Xin Li wrote:
>>> +    /*
>>> +     * If a feature is disabled during compile time clear its feature
>>> +     * bit to prevent it from showing up in the x86_capability bitmask.
>>> +     */
>>> +    if (!cpu_feature_enabled(X86_FEATURE_LAM))
>>> +        setup_clear_cpu_cap(X86_FEATURE_LAM);
>>> +
>>> +    if (!cpu_feature_enabled(X86_FEATURE_FRED))
>>> +        setup_clear_cpu_cap(X86_FEATURE_FRED);
>>> +
>> 
>> The following code will work as a generic fix:
>> 
>>     c->x86_capability[i] &= ~DISABLED_MASK(i);
>> 
>> And DISABLED_MASK(x) needs to be defined like DISABLED_MASK_BIT_SET(x).
>
>Maciej, I would much rather have a generic fix than force everyone to
>remember to open-code this for every new feature that gets disabled.

I'll apply Xin's suggestions, test and resend.

-- 
Kind regards
Maciej Wieczór-Retman