Re: [kvm-unit-tests PATCH v2 2/5] s390x: lib: fix pgtable.h

[Janosch Frank <frankja@linux.ibm.com>]

On 2/23/21 4:21 PM, Claudio Imbrenda wrote:
> On Tue, 23 Feb 2021 15:31:06 +0100
> Janosch Frank <frankja@linux.ibm.com> wrote:
> 
>> On 2/23/21 3:07 PM, Claudio Imbrenda wrote:
>>> Fix pgtable.h:
>>>
>>> * SEGMENT_ENTRY_SFAA had one extra bit set
>>> * pmd entries don't have a length
>>> * ipte does not need to clear the lower bits
>>> * pud entries should use SEGMENT_TABLE_LENGTH, as they point to
>>> segment tables
>>>
>>> Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
>>> ---
>>>  lib/s390x/asm/pgtable.h | 9 ++++-----
>>>  1 file changed, 4 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/lib/s390x/asm/pgtable.h b/lib/s390x/asm/pgtable.h
>>> index 277f3480..a2ff2d4e 100644
>>> --- a/lib/s390x/asm/pgtable.h
>>> +++ b/lib/s390x/asm/pgtable.h
>>> @@ -60,7 +60,7 @@
>>>  #define SEGMENT_SHIFT			20
>>>  
>>>  #define SEGMENT_ENTRY_ORIGIN		0xfffffffffffff800UL
>>> -#define SEGMENT_ENTRY_SFAA		0xfffffffffff80000UL
>>> +#define SEGMENT_ENTRY_SFAA		0xfffffffffff00000UL
>>>  #define SEGMENT_ENTRY_AV		0x0000000000010000UL
>>>  #define SEGMENT_ENTRY_ACC		0x000000000000f000UL
>>>  #define SEGMENT_ENTRY_F			0x0000000000000800UL
>>> @@ -183,7 +183,7 @@ static inline pmd_t *pmd_alloc(pud_t *pud,
>>> unsigned long addr) if (pud_none(*pud)) {
>>>  		pmd_t *pmd = pmd_alloc_one();
>>>  		pud_val(*pud) = __pa(pmd) |
>>> REGION_ENTRY_TT_REGION3 |
>>> -				REGION_TABLE_LENGTH;
>>> +				SEGMENT_TABLE_LENGTH;  
>>
>> @David: I'd much rather have REGION_ENTRY_LENGTH instead of
>> REGION_TABLE_LENGTH and SEGMENT_TABLE_LENGTH.
> 
> I'm weakly against it
> 
>> My argument is that this is not really an attribute of the table and
> 
> it actually is an attribute of the table. the *_TABLE_LENGTH fields
> tell how long the _table pointed to_ is. we always allocate the full 4
> pages, so in our case it will always be 0x3.
> 

It's part of the entry nevertheless and should not be a *_TABLE_*
constant. It's used in entries and has a very specific format that only
makes sense if it's being used in a region entry.

Every other thing that you or into the entry apart from the address is
named *_ENTRY_* so this should be too.

> segment table entries don't have a length because they point to page
> tables. region3 table entries point to segment tables, so they have
> SEGMENT_TABLE_LENGTH in their length field.

Btw. I'd guess the SEGMENT_TABLE_LENGTH name is the reason that you had
to fix the problem in the next hunk...

>> very much specific to the format of the (region table) entry. We
>> already have the table order as a length anyway...
>>
>> Could you tell me what you had in mind when splitting this?
>>
>>>  	}
>>>  	return pmd_offset(pud, addr);
>>>  }
>>> @@ -202,15 +202,14 @@ static inline pte_t *pte_alloc(pmd_t *pmd,
>>> unsigned long addr) {
>>>  	if (pmd_none(*pmd)) {
>>>  		pte_t *pte = pte_alloc_one();
>>> -		pmd_val(*pmd) = __pa(pte) |
>>> SEGMENT_ENTRY_TT_SEGMENT |
>>> -				SEGMENT_TABLE_LENGTH;
>>> +		pmd_val(*pmd) = __pa(pte) |
>>> SEGMENT_ENTRY_TT_SEGMENT;  
>>
>> Uhhhh good catch!
>>
>>>  	}
>>>  	return pte_offset(pmd, addr);
>>>  }
>>>  
>>>  static inline void ipte(unsigned long vaddr, pteval_t *p_pte)
>>>  {
>>> -	unsigned long table_origin = (unsigned long)p_pte &
>>> PAGE_MASK;
>>> +	unsigned long table_origin = (unsigned long)p_pte;
>>>  
>>>  	asm volatile(
>>>  		"	ipte %0,%1\n"
>>>   
>>
>> IPTE ignores that data but having the page mask also doesn't hurt so
>> generally this is not a fix, right?
> 
> apart from avoiding an unnecessary operation, there is a small nit:
> IPTE wants the address of the page table, ignoring the rightmost _11_
> bits. with PAGE_MASK we ignore the rightmost _12_ bits instead.
> it's not an issue in practice, because we allocate one page table per
> page anyway, wasting the second half of the page, so in our case that
> stray bit will always be 0. but in case we decide to allocate the page
> tables more tightly, or in case some testcase wants to manually play
> tricks with the page tables, there might be the risk that IPTE would
> target the wrong page table.
> 
> by not clearing the lower 12 bits we not only save an unnecessary
> operation, but we are actually more architecturally correct.
> 

Right, the old 2k pgtable thing...
Would you mind putting that into the patch description?