Re: [PATCH v3 09/33] nds32: Cache and TLB routines

[Greentime Hu]

2017-12-13 16:53 GMT+08:00 Guo Ren <ren_guo@xxxxxxxxx>:
> On Wed, Dec 13, 2017 at 04:30:41PM +0800, Greentime Hu wrote:
>> 2017-12-13 16:19 GMT+08:00 Guo Ren <ren_guo@xxxxxxxxx>:
>> > On Wed, Dec 13, 2017 at 01:45:02PM +0800, Greentime Hu wrote:
>> >
>> >> I think it should be fine if an interruption between mtsr_dsb and
>> >> tlbop_rwr because this is a optimization by sw.
>> >
>> > Fine? When there is an unexpected vaddr in SR_TLB_VPN, tlbop_rwr(*pte) will
>> > break that vaddr's pfn in the CPU tlb-buffer entry. When linux access the
>> > vaddr, it will get wrong data unless the entry has been replaced out.
>>
>> Hi, Guo Ren:
>>
>> Thanks. I get your point.
>> It is needed to be protected.
>> I will fix it in the next version patch.
>>
>> if (vma->vm_mm == current->active_mm) {
>>         local_irq_save(flags);
>>         __nds32__mtsr_dsb(addr, NDS32_SR_TLB_VPN);
>>         __nds32__tlbop_rwr(*pte);
>>         __nds32__isb();
>>         local_irq_restore(flags);
>> }
>
> If hardware tlbop_rwr could invalid NDS32_SR_TLB_VPN, then you needn't
> protect.
> I mean:
> mtsr addr1 NDS32_SR_TLB_VPN
> mtsr addr2 NDS32_SR_TLB_VPN
> tlbop_rwr(*pte) // OK, and it will hit a hardware invalid bit internal.
> tlbop_rwr(*pte) // SR_TLB_VPN invalided, then it will not cause problem.
>
> :) How my idea?

Hi, Guo Ren:

The reason I think it might have problem is that

mtsr addr1 NDS32_SR_TLB_VPN
    interrupt coming
        mtsr addr2 NDS32_SR_TLB_VPN <- TLB_VPN has been set to addr2
        tlbop_rwr(*pte);
    interrupt finish
tlbop_rwr(*pte); <- it will use the wrong TLB_VPN

I think all these TLB operations should be protected because tlbop
will operate with TLB_VPN.
Thanks :)