Re: [RFC 06/10] arm64/mm: Add FEAT_LPA2 specific encoding

[Anshuman Khandual]

On 7/16/21 3:32 PM, Steven Price wrote:
> On 16/07/2021 08:20, Anshuman Khandual wrote:
>>
>>
>> On 7/14/21 9:08 PM, Steven Price wrote:
>>> On 14/07/2021 03:21, Anshuman Khandual wrote:
>>>> FEAT_LPA2 requires different PTE representation formats for both 4K and 16K
>>>> page size config. This adds FEAT_LPA2 specific new PTE encodings as per ARM
>>>> ARM (0487G.A) which updates [pte|phys]_to_[phys|pte](). The updated helpers
>>>> would be used when FEAT_LPA2 gets enabled via CONFIG_ARM64_PA_BITS_52 on 4K
>>>> and 16K page size. Although TTBR encoding and phys_to_ttbr() helper remains
>>>> the same as FEAT_LPA for FEAT_LPA2 as well. It updates 'phys_to_pte' helper
>>>> to accept a temporary variable and changes impacted call sites.
>>>>
>>>> Signed-off-by: Anshuman Khandual <anshuman.khandual@xxxxxxx>
>>>> ---
>>>>  arch/arm64/include/asm/assembler.h     | 23 +++++++++++++++++++----
>>>>  arch/arm64/include/asm/pgtable-hwdef.h |  4 ++++
>>>>  arch/arm64/include/asm/pgtable.h       |  4 ++++
>>>>  arch/arm64/kernel/head.S               | 25 +++++++++++++------------
>>>>  4 files changed, 40 insertions(+), 16 deletions(-)
>>>>
>>>> diff --git a/arch/arm64/include/asm/assembler.h b/arch/arm64/include/asm/assembler.h
>>>> index fedc202..0492543 100644
>>>> --- a/arch/arm64/include/asm/assembler.h
>>>> +++ b/arch/arm64/include/asm/assembler.h
>>>> @@ -606,7 +606,7 @@ alternative_endif
>>>>  #endif
>>>>  	.endm
>>>>  
>>>> -	.macro	phys_to_pte, pte, phys
>>>> +	.macro	phys_to_pte, pte, phys, tmp
>>>>  #ifdef CONFIG_ARM64_PA_BITS_52_LPA
>>>>  	/*
>>>>  	 * We assume \phys is 64K aligned and this is guaranteed by only
>>>> @@ -614,6 +614,17 @@ alternative_endif
>>>>  	 */
>>>>  	orr	\pte, \phys, \phys, lsr #36
>>>>  	and	\pte, \pte, #PTE_ADDR_MASK
>>>> +#elif defined(CONFIG_ARM64_PA_BITS_52_LPA2)
>>>> +	orr	\pte, \phys, \phys, lsr #42
>>>> +
>>>> +	/*
>>>> +	 * The 'tmp' is being used here to just prepare
>>>> +	 * and hold PTE_ADDR_MASK which cannot be passed
>>>> +	 * to the subsequent 'and' instruction.
>>>> +	 */
>>>> +	mov	\tmp, #PTE_ADDR_LOW
>>>> +	orr	\tmp, \tmp, #PTE_ADDR_HIGH
>>>> +	and	\pte, \pte, \tmp
>>> Rather than adding an extra temporary register (and the fallout of
>>> various other macros needing an extra register), this can be done with
>>> two AND instructions:
>>
>> I would really like to get rid of the 'tmp' variable here as
>> well but did not figure out any method of accomplishing it.
>>
>>>
>>> 	/* PTE_ADDR_MASK cannot be encoded as an immediate, so
>>>          * mask off all but two bits, followed by masking the
>>>          * extra two bits
>>>          */
>>> 	and	\pte, \pte, #PTE_ADDR_MASK | (3 << 10)
>>> 	and	\pte, \pte, #~(3 << 10)
>>
>> Did this change as suggested
>>
>> --- a/arch/arm64/include/asm/assembler.h
>> +++ b/arch/arm64/include/asm/assembler.h
>> @@ -626,9 +626,8 @@ alternative_endif
>>          * and hold PTE_ADDR_MASK which cannot be passed
>>          * to the subsequent 'and' instruction.
>>          */
>> -       mov     \tmp, #PTE_ADDR_LOW
>> -       orr     \tmp, \tmp, #PTE_ADDR_HIGH
>> -       and     \pte, \pte, \tmp
>> +       and     \pte, \pte, #PTE_ADDR_MASK | (0x3 << 10)
>> +       and     \pte, \pte, #~(0x3 << 10)
>>  
>>  .Lskip_lpa2\@:
>>         mov     \pte, \phys
>>
>>
>> but still fails to build (tested on 16K)
>>
>> arch/arm64/kernel/head.S: Assembler messages:
>> arch/arm64/kernel/head.S:377: Error: immediate out of range at operand 3 -- `and x6,x6,#((((1<<(50-14))-1)<<14)|(0x3<<8))|(0x3<<10)'
>> arch/arm64/kernel/head.S:390: Error: immediate out of range at operand 3 -- `and x12,x12,#((((1<<(50-14))-1)<<14)|(0x3<<8))|(0x3<<10)'
>> arch/arm64/kernel/head.S:390: Error: immediate out of range at operand 3 -- `and x12,x12,#((((1<<(50-14))-1)<<14)|(0x3<<8))|(0x3<<10)'
>> arch/arm64/kernel/head.S:404: Error: immediate out of range at operand 3 -- `and x12,x12,#((((1<<(50-14))-1)<<14)|(0x3<<8))|(0x3<<10)'
>> arch/arm64/kernel/head.S:404: Error: immediate out of range at operand 3 -- `and x12,x12,#((((1<<(50-14))-1)<<14)|(0x3<<8))|(0x3<<10)'
>>
> 
> Ah, I'd only tested this for 4k. 16k would require a different set of masks.
> 
> So the bits we need to cover are those from just below PAGE_SHIFT to the
> top of PTE_ADDR_HIGH (bit 10). So we can compute the mask for both 4k

Okay.

> and 16k with GENMASK(PAGE_SHIFT-1, 10):
> 
> 	and	\pte, \pte, #PTE_ADDR_MASK | GENMASK(PAGE_SHIFT - 1, 10)
> 	and	\pte, \pte, #~GENMASK(PAGE_SHIFT - 1, 10)
> 
> This compiles (for both 4k and 16k) and the assembly looks correct, but
> I've not done any other testing.

Yeah it works, will do the change.