Re: tlb_batch_add_one()

[Nitin Gupta <nitin.m.gupta@oracle.com>]

On 3/30/17 7:57 PM, David Miller wrote:
> From: Nitin Gupta <nitin.m.gupta@oracle.com>
> Date: Thu, 30 Mar 2017 17:42:51 -0700
> 
>> diff --git a/arch/sparc/mm/tsb.c b/arch/sparc/mm/tsb.c
>> index 0a04811..bedf08b 100644
>> --- a/arch/sparc/mm/tsb.c
>> +++ b/arch/sparc/mm/tsb.c
>> @@ -122,7 +122,7 @@ void flush_tsb_user(struct tlb_batch *tb)
>>
>>  	spin_lock_irqsave(&mm->context.lock, flags);
>>
>> -	if (tb->hugepage_shift < HPAGE_SHIFT) {
>> +	if (tb->hugepage_shift < REAL_HPAGE_SHIFT) {
>>  		base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
>>  		nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
>>  		if (tlb_type = cheetah_plus || tlb_type = hypervisor)
>> @@ -155,7 +155,7 @@ void flush_tsb_user_page(struct mm_struct *mm,
>> unsigned long vaddr,
>>
>>  	spin_lock_irqsave(&mm->context.lock, flags);
>>
>> -	if (hugepage_shift < HPAGE_SHIFT) {
>> +	if (hugepage_shift < REAL_HPAGE_SHIFT) {
>>  		base = (unsigned long) mm->context.tsb_block[MM_TSB_BASE].tsb;
>>  		nentries = mm->context.tsb_block[MM_TSB_BASE].tsb_nentries;
>>  		if (tlb_type = cheetah_plus || tlb_type = hypervisor)
>> -- 
> 
> I think if we do it like this, it will only flush one half of the huge
> page.
>

Flushing only half the 8M page is the intended behavior: after the
initial allocation of 8M hugepage, the page is handled exactly as if two
independent 4M hugepages were allocated (that happen to be physically
contiguous). So, for each 4M chunk, flushing from TLB and TSB is done
independently. For instance, in set_huge_pte_at() we added special case
for (size = HPAGE_SIZE) to flush the "second half" of 8M page.
Similarly in huge_ptep_get_and_clear() and in set_pmd_at().

Nitin


> We really need to pass HPAGE_SHIFT down into here, so that the TSB
> flush gets both REAL_HPAGE_SIZE entries.
>