Re: [PATCH v3 2/8] x86/mm: Change the leave_mm() condition for local TLB flushes

[Andy Lutomirski]

On Thu, May 25, 2017 at 6:39 PM, Rik van Riel <riel@xxxxxxxxxx> wrote:
> On Thu, 2017-05-25 at 17:47 -0700, Andy Lutomirski wrote:
>>
>> +++ b/arch/x86/mm/tlb.c
>> @@ -311,7 +311,7 @@ void flush_tlb_mm_range(struct mm_struct *mm,
>> unsigned long start,
>>               goto out;
>>       }
>>
>> -     if (!current->mm) {
>> +     if (this_cpu_read(cpu_tlbstate.state) != TLBSTATE_OK) {
>>               leave_mm(smp_processor_id());
>
> Unless -mm changed leave_mm (I did not check), this
> is not quite correct yet.
>
> The reason is leave_mm (at least in the latest Linus
> tree) ignores the cpu argument for one of its checks.
>
> You should probably fix that in an earlier patch,
> assuming you haven't already done so in -mm.
>
> void leave_mm(int cpu)
> {
>         struct mm_struct *active_mm =
> this_cpu_read(cpu_tlbstate.active_mm);
>         if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK)
>                 BUG();
>         if (cpumask_test_cpu(cpu, mm_cpumask(active_mm))) {
>                 cpumask_clear_cpu(cpu, mm_cpumask(active_mm));
>                 load_cr3(swapper_pg_dir);

I agree it's odd, but what's the bug?  Both before and after, leave_mm
needed to be called with cpu == smp_processor_id(), and
smp_processor_id() warns if it's called in a preemptible context.