[Devel] [PATCH vz8 v2 1/2] x86, cpuinfo: Fix race on parallel /proc/cpuinfo read

[Andrey Ryabinin]

On 11/3/20 2:28 PM, Kirill Tkhai wrote:
> On 02.11.2020 20:13, Andrey Ryabinin wrote:
>> If several threads read /proc/cpuinfo some can see in 'flags'
>> values from c->x86_capability, before __do_cpuid_fault() called
>> and masks applied. Fix this by forming 'flags' on stack first
>> and copy them in per_cpu(cpu_flags, cpu) as a last step.
>>
>> https://jira.sw.ru/browse/PSBM-121823
>> Signed-off-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
>> ---
>> Changes since v1:
>>  - none
>>  
>>  arch/x86/kernel/cpu/proc.c | 17 +++++++++--------
>>  1 file changed, 9 insertions(+), 8 deletions(-)
>>
>> diff --git a/arch/x86/kernel/cpu/proc.c b/arch/x86/kernel/cpu/proc.c
>> index 4fe1577d5e6f..4cc2951e34fb 100644
>> --- a/arch/x86/kernel/cpu/proc.c
>> +++ b/arch/x86/kernel/cpu/proc.c
>> @@ -69,11 +69,11 @@ static DEFINE_PER_CPU(struct cpu_flags, cpu_flags);
>>  static void init_cpu_flags(void *dummy)
>>  {
>>  	int cpu = smp_processor_id();
>> -	struct cpu_flags *flags = &per_cpu(cpu_flags, cpu);
>> +	struct cpu_flags flags;
>>  	struct cpuinfo_x86 *c = &cpu_data(cpu);
>>  	unsigned int eax, ebx, ecx, edx;
>>  
>> -	memcpy(flags->val, c->x86_capability, NCAPINTS * sizeof(u32));
>> +	memcpy(&flags, c->x86_capability, sizeof(flags));
>>  
>>  	/*
>>  	 * Clear feature bits masked using cpuid masking/faulting.
>> @@ -81,26 +81,27 @@ static void init_cpu_flags(void *dummy)
>>  
>>  	if (c->cpuid_level >= 0x00000001) {
>>  		__do_cpuid_fault(0x00000001, 0, &eax, &ebx, &ecx, &edx);
>> -		flags->val[4] &= ecx;
>> -		flags->val[0] &= edx;
>> +		flags.val[4] &= ecx;
>> +		flags.val[0] &= edx;
>>  	}
>>  
>>  	if (c->cpuid_level >= 0x00000007) {
>>  		__do_cpuid_fault(0x00000007, 0, &eax, &ebx, &ecx, &edx);
>> -		flags->val[9] &= ebx;
>> +		flags.val[9] &= ebx;
>>  	}
>>  
>>  	if ((c->extended_cpuid_level & 0xffff0000) == 0x80000000 &&
>>  	    c->extended_cpuid_level >= 0x80000001) {
>>  		__do_cpuid_fault(0x80000001, 0, &eax, &ebx, &ecx, &edx);
>> -		flags->val[6] &= ecx;
>> -		flags->val[1] &= edx;
>> +		flags.val[6] &= ecx;
>> +		flags.val[1] &= edx;
>>  	}
>>  
>>  	if (c->cpuid_level >= 0x0000000d) {
>>  		__do_cpuid_fault(0x0000000d, 1, &eax, &ebx, &ecx, &edx);
>> -		flags->val[10] &= eax;
>> +		flags.val[10] &= eax;
>>  	}
>> +	memcpy(&per_cpu(cpu_flags, cpu), &flags, sizeof(flags));
> 
> This is still racy, since memcpy() is not atomic. Maybe we should add some lock on top of this?
> 

This race shouldn't be a problem since flags are not supposed to change during ve lifetime.
So we overwriting same values. But don't mind to add spinlock protection.