lkml.org 
[lkml]   [2019]   [Jan]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v3 4/7] arm64: add sysfs vulnerability show for meltdown
From
Date
Hi Julien,

On 01/10/2019 03:23 AM, Julien Thierry wrote:
> Hi Jeremy,
>
> On 09/01/2019 23:55, Jeremy Linton wrote:
>> Display the mitigation status if active, otherwise
>> assume the cpu is safe unless it doesn't have CSV3
>> and isn't in our whitelist.
>>
>> Signed-off-by: Jeremy Linton <jeremy.linton@arm.com>
>> ---
>> arch/arm64/kernel/cpufeature.c | 32 +++++++++++++++++++++++++++-----
>> 1 file changed, 27 insertions(+), 5 deletions(-)
>>
>> diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c
>> index ab784d7a0083..ef7bbc49ef78 100644
>> --- a/arch/arm64/kernel/cpufeature.c
>> +++ b/arch/arm64/kernel/cpufeature.c
>> @@ -944,8 +944,12 @@ has_useable_cnp(const struct arm64_cpu_capabilities *entry, int scope)
>> return has_cpuid_feature(entry, scope);
>> }
>>
>> +/* default value is invalid until unmap_kernel_at_el0() runs */
>> +static bool __meltdown_safe = true;
>> +
>> #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
>> static int __kpti_forced; /* 0: not forced, >0: forced on, <0: forced off */
>> +extern uint arm64_requested_vuln_attrs;
>>
>> static bool is_cpu_meltdown_safe(void)
>> {
>> @@ -972,6 +976,14 @@ static bool unmap_kernel_at_el0(const struct arm64_cpu_capabilities *entry,
>> {
>> char const *str = "command line option";
>>
>> + bool meltdown_safe = is_cpu_meltdown_safe() ||
>> + has_cpuid_feature(entry, scope);
>> +
>> + if (!meltdown_safe)
>> + __meltdown_safe = false;
>> +
>> + arm64_requested_vuln_attrs |= VULN_MELTDOWN;
>> +
>> /*
>> * For reasons that aren't entirely clear, enabling KPTI on Cavium
>> * ThunderX leads to apparent I-cache corruption of kernel text, which
>> @@ -993,11 +1005,7 @@ static bool unmap_kernel_at_el0(const struct arm64_cpu_capabilities *entry,
>> if (IS_ENABLED(CONFIG_RANDOMIZE_BASE))
>> return true;
>>
>> - if (is_cpu_meltdown_safe())
>> - return false;
>> -
>> - /* Defer to CPU feature registers */
>> - return !has_cpuid_feature(entry, scope);
>> + return !meltdown_safe;
>> }
>>
>> static void
>> @@ -2065,3 +2073,17 @@ static int __init enable_mrs_emulation(void)
>> }
>>
>> core_initcall(enable_mrs_emulation);
>> +
>> +#ifdef CONFIG_GENERIC_CPU_VULNERABILITIES
>> +ssize_t cpu_show_meltdown(struct device *dev, struct device_attribute *attr,
>> + char *buf)
>> +{
>> + if (arm64_kernel_unmapped_at_el0())
>> + return sprintf(buf, "Mitigation: KPTI\n");
>> +
>> + if (__meltdown_safe)
>> + return sprintf(buf, "Not affected\n");
>
> An issue I see is that we don't even bother to check it that CPUs are
> meltdown safe if CONFIG_UNMAP_KERNEL_AT_EL0 is not defined but here
> we'll advertise that the system is meltdown safe.

That check isn't necessary anymore because the sysfs attribute is only
populated if unmap_kernel_at_el0() runs (assuming I haven't messed
something up). That was Dave/Will's suggestions in the last thread about
how to handle this case.



>
> I think that checking whether we know that CPUs are meltdown safe should
> be separated from whether mitigation is applied.
>
> Someone who knows thinks their CPUs are in the white list might want to
> compile out code that does the kpti, but it would be good to give them a
> proper diagnostic whether they were wrong or not.
>
> Cheers,
>

\
 
 \ /
  Last update: 2019-01-10 15:11    [W:0.078 / U:3.328 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site