diff mbox series

[v2,2/2] ARM: support PHYS_OFFSET minimum aligned at 64KiB boundary

Message ID 20200915131615.3138-3-thunder.leizhen@huawei.com (mailing list archive)
State New, archived
Headers show
Series ARM: support PHYS_OFFSET minimum aligned at 64KiB boundary | expand

Commit Message

Leizhen (ThunderTown) Sept. 15, 2020, 1:16 p.m. UTC
Currently, only support the kernels where the base of physical memory is
at a 16MiB boundary. Because the add/sub instructions only contains 8bits
unrotated value. But we can use one more "add/sub" instructions to handle
bits 23-16. The performance will be slightly affected.

Since most boards meet 16 MiB alignment, so add a new configuration
option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if
anyone really needs it.

All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are
used in __fixup_a_pv_table() now, but the callee saved r11 is not used in
the whole head.S file. So choose it.

Because the calculation of "y = x + __pv_offset[63:24]" have been done,
so we only need to calculate "y = y + __pv_offset[23:16]", that's why
the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub()
in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t"
(above y).

Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
---
 arch/arm/Kconfig              | 18 +++++++++++++++++-
 arch/arm/include/asm/memory.h | 16 +++++++++++++---
 arch/arm/kernel/head.S        | 25 +++++++++++++++++++------
 3 files changed, 49 insertions(+), 10 deletions(-)

Comments

Russell King (Oracle) Sept. 15, 2020, 7:01 p.m. UTC | #1
On Tue, Sep 15, 2020 at 09:16:15PM +0800, Zhen Lei wrote:
> Currently, only support the kernels where the base of physical memory is
> at a 16MiB boundary. Because the add/sub instructions only contains 8bits
> unrotated value. But we can use one more "add/sub" instructions to handle
> bits 23-16. The performance will be slightly affected.
> 
> Since most boards meet 16 MiB alignment, so add a new configuration
> option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if
> anyone really needs it.
> 
> All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are
> used in __fixup_a_pv_table() now, but the callee saved r11 is not used in
> the whole head.S file. So choose it.
> 
> Because the calculation of "y = x + __pv_offset[63:24]" have been done,
> so we only need to calculate "y = y + __pv_offset[23:16]", that's why
> the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub()
> in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t"
> (above y).
> 
> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
> ---
>  arch/arm/Kconfig              | 18 +++++++++++++++++-
>  arch/arm/include/asm/memory.h | 16 +++++++++++++---
>  arch/arm/kernel/head.S        | 25 +++++++++++++++++++------
>  3 files changed, 49 insertions(+), 10 deletions(-)
> 
> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
> index e00d94b16658765..19fc2c746e2ce29 100644
> --- a/arch/arm/Kconfig
> +++ b/arch/arm/Kconfig
> @@ -240,12 +240,28 @@ config ARM_PATCH_PHYS_VIRT
>  	  kernel in system memory.
>  
>  	  This can only be used with non-XIP MMU kernels where the base
> -	  of physical memory is at a 16MB boundary.
> +	  of physical memory is at a 16MiB boundary.
>  
>  	  Only disable this option if you know that you do not require
>  	  this feature (eg, building a kernel for a single machine) and
>  	  you need to shrink the kernel to the minimal size.
>  
> +config ARM_PATCH_PHYS_VIRT_RADICAL
> +	bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary"
> +	default n

Please drop the "default n" - this is the default anyway.

> @@ -236,6 +243,9 @@ static inline unsigned long __phys_to_virt(phys_addr_t x)
>  	 * in place where 'r' 32 bit operand is expected.
>  	 */
>  	__pv_stub((unsigned long) x, t, "sub", __PV_BITS_31_24);
> +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
> +	__pv_stub((unsigned long) t, t, "sub", __PV_BITS_23_16);

t is already unsigned long, so this cast is not necessary.

I've been debating whether it would be better to use "movw" for this
for ARMv7.  In other words:

	movw	tmp, #16-bit
	adds	%Q0, %1, tmp, lsl #16
	adc	%R0, %R0, #0

It would certainly be less instructions, but at the cost of an
additional register - and we'd have to change the fixup code to
know about movw.

Thoughts?
Leizhen (ThunderTown) Sept. 16, 2020, 1:57 a.m. UTC | #2
On 2020/9/16 3:01, Russell King - ARM Linux admin wrote:
> On Tue, Sep 15, 2020 at 09:16:15PM +0800, Zhen Lei wrote:
>> Currently, only support the kernels where the base of physical memory is
>> at a 16MiB boundary. Because the add/sub instructions only contains 8bits
>> unrotated value. But we can use one more "add/sub" instructions to handle
>> bits 23-16. The performance will be slightly affected.
>>
>> Since most boards meet 16 MiB alignment, so add a new configuration
>> option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if
>> anyone really needs it.
>>
>> All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are
>> used in __fixup_a_pv_table() now, but the callee saved r11 is not used in
>> the whole head.S file. So choose it.
>>
>> Because the calculation of "y = x + __pv_offset[63:24]" have been done,
>> so we only need to calculate "y = y + __pv_offset[23:16]", that's why
>> the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub()
>> in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t"
>> (above y).
>>
>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>> ---
>>  arch/arm/Kconfig              | 18 +++++++++++++++++-
>>  arch/arm/include/asm/memory.h | 16 +++++++++++++---
>>  arch/arm/kernel/head.S        | 25 +++++++++++++++++++------
>>  3 files changed, 49 insertions(+), 10 deletions(-)
>>
>> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
>> index e00d94b16658765..19fc2c746e2ce29 100644
>> --- a/arch/arm/Kconfig
>> +++ b/arch/arm/Kconfig
>> @@ -240,12 +240,28 @@ config ARM_PATCH_PHYS_VIRT
>>  	  kernel in system memory.
>>  
>>  	  This can only be used with non-XIP MMU kernels where the base
>> -	  of physical memory is at a 16MB boundary.
>> +	  of physical memory is at a 16MiB boundary.
>>  
>>  	  Only disable this option if you know that you do not require
>>  	  this feature (eg, building a kernel for a single machine) and
>>  	  you need to shrink the kernel to the minimal size.
>>  
>> +config ARM_PATCH_PHYS_VIRT_RADICAL
>> +	bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary"
>> +	default n
> 
> Please drop the "default n" - this is the default anyway.

OK, I will remove it.

> 
>> @@ -236,6 +243,9 @@ static inline unsigned long __phys_to_virt(phys_addr_t x)
>>  	 * in place where 'r' 32 bit operand is expected.
>>  	 */
>>  	__pv_stub((unsigned long) x, t, "sub", __PV_BITS_31_24);
>> +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
>> +	__pv_stub((unsigned long) t, t, "sub", __PV_BITS_23_16);
> 
> t is already unsigned long, so this cast is not necessary.

Oh, yes, yes, I copied from the above statement, but forgot to remove it.

> 
> I've been debating whether it would be better to use "movw" for this
> for ARMv7.  In other words:
> 
> 	movw	tmp, #16-bit
> 	adds	%Q0, %1, tmp, lsl #16
> 	adc	%R0, %R0, #0
> 
> It would certainly be less instructions, but at the cost of an
> additional register - and we'd have to change the fixup code to
> know about movw.

It's one less instruction for 64KiB boundary && (sizeof(phys_addr_t) == 8),
and no increase or decrease for 64KiB boundary && (sizeof(phys_addr_t) == 4),
but one more instruction for 16MiB boundary.

And maybe: 16MiB is widely used, but 64KiB is rarely used.

So I'm inclined to the current revision.

> 
> Thoughts?
>
Russell King (Oracle) Sept. 16, 2020, 7:57 a.m. UTC | #3
On Wed, Sep 16, 2020 at 09:57:15AM +0800, Leizhen (ThunderTown) wrote:
> On 2020/9/16 3:01, Russell King - ARM Linux admin wrote:
> > On Tue, Sep 15, 2020 at 09:16:15PM +0800, Zhen Lei wrote:
> >> Currently, only support the kernels where the base of physical memory is
> >> at a 16MiB boundary. Because the add/sub instructions only contains 8bits
> >> unrotated value. But we can use one more "add/sub" instructions to handle
> >> bits 23-16. The performance will be slightly affected.
> >>
> >> Since most boards meet 16 MiB alignment, so add a new configuration
> >> option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if
> >> anyone really needs it.
> >>
> >> All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are
> >> used in __fixup_a_pv_table() now, but the callee saved r11 is not used in
> >> the whole head.S file. So choose it.
> >>
> >> Because the calculation of "y = x + __pv_offset[63:24]" have been done,
> >> so we only need to calculate "y = y + __pv_offset[23:16]", that's why
> >> the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub()
> >> in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t"
> >> (above y).
> >>
> >> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
> >> ---
> >>  arch/arm/Kconfig              | 18 +++++++++++++++++-
> >>  arch/arm/include/asm/memory.h | 16 +++++++++++++---
> >>  arch/arm/kernel/head.S        | 25 +++++++++++++++++++------
> >>  3 files changed, 49 insertions(+), 10 deletions(-)
> >>
> >> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
> >> index e00d94b16658765..19fc2c746e2ce29 100644
> >> --- a/arch/arm/Kconfig
> >> +++ b/arch/arm/Kconfig
> >> @@ -240,12 +240,28 @@ config ARM_PATCH_PHYS_VIRT
> >>  	  kernel in system memory.
> >>  
> >>  	  This can only be used with non-XIP MMU kernels where the base
> >> -	  of physical memory is at a 16MB boundary.
> >> +	  of physical memory is at a 16MiB boundary.
> >>  
> >>  	  Only disable this option if you know that you do not require
> >>  	  this feature (eg, building a kernel for a single machine) and
> >>  	  you need to shrink the kernel to the minimal size.
> >>  
> >> +config ARM_PATCH_PHYS_VIRT_RADICAL
> >> +	bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary"
> >> +	default n
> > 
> > Please drop the "default n" - this is the default anyway.
> 
> OK, I will remove it.
> 
> > 
> >> @@ -236,6 +243,9 @@ static inline unsigned long __phys_to_virt(phys_addr_t x)
> >>  	 * in place where 'r' 32 bit operand is expected.
> >>  	 */
> >>  	__pv_stub((unsigned long) x, t, "sub", __PV_BITS_31_24);
> >> +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
> >> +	__pv_stub((unsigned long) t, t, "sub", __PV_BITS_23_16);
> > 
> > t is already unsigned long, so this cast is not necessary.
> 
> Oh, yes, yes, I copied from the above statement, but forgot to remove it.
> 
> > 
> > I've been debating whether it would be better to use "movw" for this
> > for ARMv7.  In other words:
> > 
> > 	movw	tmp, #16-bit
> > 	adds	%Q0, %1, tmp, lsl #16
> > 	adc	%R0, %R0, #0
> > 
> > It would certainly be less instructions, but at the cost of an
> > additional register - and we'd have to change the fixup code to
> > know about movw.
> 
> It's one less instruction for 64KiB boundary && (sizeof(phys_addr_t) == 8),
> and no increase or decrease for 64KiB boundary && (sizeof(phys_addr_t) == 4),
> but one more instruction for 16MiB boundary.
> 
> And maybe: 16MiB is widely used, but 64KiB is rarely used.
> 
> So I'm inclined to the current revision.

Multiplatform kernels (which will be what distros build) will have to
enable this option if they wish to support this platform. So, in that
case it doesn't just impacting a single platform, but all platforms.
Leizhen (ThunderTown) Sept. 17, 2020, 3:26 a.m. UTC | #4
On 2020/9/16 15:57, Russell King - ARM Linux admin wrote:
> On Wed, Sep 16, 2020 at 09:57:15AM +0800, Leizhen (ThunderTown) wrote:
>> On 2020/9/16 3:01, Russell King - ARM Linux admin wrote:
>>> On Tue, Sep 15, 2020 at 09:16:15PM +0800, Zhen Lei wrote:
>>>> Currently, only support the kernels where the base of physical memory is
>>>> at a 16MiB boundary. Because the add/sub instructions only contains 8bits
>>>> unrotated value. But we can use one more "add/sub" instructions to handle
>>>> bits 23-16. The performance will be slightly affected.
>>>>
>>>> Since most boards meet 16 MiB alignment, so add a new configuration
>>>> option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if
>>>> anyone really needs it.
>>>>
>>>> All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are
>>>> used in __fixup_a_pv_table() now, but the callee saved r11 is not used in
>>>> the whole head.S file. So choose it.
>>>>
>>>> Because the calculation of "y = x + __pv_offset[63:24]" have been done,
>>>> so we only need to calculate "y = y + __pv_offset[23:16]", that's why
>>>> the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub()
>>>> in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t"
>>>> (above y).
>>>>
>>>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>>>> ---
>>>>  arch/arm/Kconfig              | 18 +++++++++++++++++-
>>>>  arch/arm/include/asm/memory.h | 16 +++++++++++++---
>>>>  arch/arm/kernel/head.S        | 25 +++++++++++++++++++------
>>>>  3 files changed, 49 insertions(+), 10 deletions(-)
>>>>
>>>> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
>>>> index e00d94b16658765..19fc2c746e2ce29 100644
>>>> --- a/arch/arm/Kconfig
>>>> +++ b/arch/arm/Kconfig
>>>> @@ -240,12 +240,28 @@ config ARM_PATCH_PHYS_VIRT
>>>>  	  kernel in system memory.
>>>>  
>>>>  	  This can only be used with non-XIP MMU kernels where the base
>>>> -	  of physical memory is at a 16MB boundary.
>>>> +	  of physical memory is at a 16MiB boundary.
>>>>  
>>>>  	  Only disable this option if you know that you do not require
>>>>  	  this feature (eg, building a kernel for a single machine) and
>>>>  	  you need to shrink the kernel to the minimal size.
>>>>  
>>>> +config ARM_PATCH_PHYS_VIRT_RADICAL
>>>> +	bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary"
>>>> +	default n
>>>
>>> Please drop the "default n" - this is the default anyway.
>>
>> OK, I will remove it.
>>
>>>
>>>> @@ -236,6 +243,9 @@ static inline unsigned long __phys_to_virt(phys_addr_t x)
>>>>  	 * in place where 'r' 32 bit operand is expected.
>>>>  	 */
>>>>  	__pv_stub((unsigned long) x, t, "sub", __PV_BITS_31_24);
>>>> +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
>>>> +	__pv_stub((unsigned long) t, t, "sub", __PV_BITS_23_16);
>>>
>>> t is already unsigned long, so this cast is not necessary.
>>
>> Oh, yes, yes, I copied from the above statement, but forgot to remove it.
>>
>>>
>>> I've been debating whether it would be better to use "movw" for this
>>> for ARMv7.  In other words:
>>>
>>> 	movw	tmp, #16-bit
>>> 	adds	%Q0, %1, tmp, lsl #16
>>> 	adc	%R0, %R0, #0
>>>
>>> It would certainly be less instructions, but at the cost of an
>>> additional register - and we'd have to change the fixup code to
>>> know about movw.
>>
>> It's one less instruction for 64KiB boundary && (sizeof(phys_addr_t) == 8),
>> and no increase or decrease for 64KiB boundary && (sizeof(phys_addr_t) == 4),
>> but one more instruction for 16MiB boundary.
>>
>> And maybe: 16MiB is widely used, but 64KiB is rarely used.
>>
>> So I'm inclined to the current revision.
> 
> Multiplatform kernels (which will be what distros build) will have to
> enable this option if they wish to support this platform. So, in that
> case it doesn't just impacting a single platform, but all platforms.

I will try movw. But it may take a few days, because I feel that the changes
will be a little big.

>
Ard Biesheuvel Sept. 17, 2020, 2 p.m. UTC | #5
On Tue, 15 Sep 2020 at 22:06, Russell King - ARM Linux admin
<linux@armlinux.org.uk> wrote:
>
> On Tue, Sep 15, 2020 at 09:16:15PM +0800, Zhen Lei wrote:
> > Currently, only support the kernels where the base of physical memory is
> > at a 16MiB boundary. Because the add/sub instructions only contains 8bits
> > unrotated value. But we can use one more "add/sub" instructions to handle
> > bits 23-16. The performance will be slightly affected.
> >
> > Since most boards meet 16 MiB alignment, so add a new configuration
> > option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if
> > anyone really needs it.
> >
> > All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are
> > used in __fixup_a_pv_table() now, but the callee saved r11 is not used in
> > the whole head.S file. So choose it.
> >
> > Because the calculation of "y = x + __pv_offset[63:24]" have been done,
> > so we only need to calculate "y = y + __pv_offset[23:16]", that's why
> > the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub()
> > in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t"
> > (above y).
> >
> > Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
> > ---
> >  arch/arm/Kconfig              | 18 +++++++++++++++++-
> >  arch/arm/include/asm/memory.h | 16 +++++++++++++---
> >  arch/arm/kernel/head.S        | 25 +++++++++++++++++++------
> >  3 files changed, 49 insertions(+), 10 deletions(-)
> >
> > diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
> > index e00d94b16658765..19fc2c746e2ce29 100644
> > --- a/arch/arm/Kconfig
> > +++ b/arch/arm/Kconfig
> > @@ -240,12 +240,28 @@ config ARM_PATCH_PHYS_VIRT
> >         kernel in system memory.
> >
> >         This can only be used with non-XIP MMU kernels where the base
> > -       of physical memory is at a 16MB boundary.
> > +       of physical memory is at a 16MiB boundary.
> >
> >         Only disable this option if you know that you do not require
> >         this feature (eg, building a kernel for a single machine) and
> >         you need to shrink the kernel to the minimal size.
> >
> > +config ARM_PATCH_PHYS_VIRT_RADICAL
> > +     bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary"
> > +     default n
>
> Please drop the "default n" - this is the default anyway.
>
> > @@ -236,6 +243,9 @@ static inline unsigned long __phys_to_virt(phys_addr_t x)
> >        * in place where 'r' 32 bit operand is expected.
> >        */
> >       __pv_stub((unsigned long) x, t, "sub", __PV_BITS_31_24);
> > +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
> > +     __pv_stub((unsigned long) t, t, "sub", __PV_BITS_23_16);
>
> t is already unsigned long, so this cast is not necessary.
>
> I've been debating whether it would be better to use "movw" for this
> for ARMv7.  In other words:
>
>         movw    tmp, #16-bit
>         adds    %Q0, %1, tmp, lsl #16
>         adc     %R0, %R0, #0
>
> It would certainly be less instructions, but at the cost of an
> additional register - and we'd have to change the fixup code to
> know about movw.
>
> Thoughts?
>

Since LPAE implies v7, we can use movw unconditionally, which is nice.

There is no need to use an additional temp register, as we can use the
register holding the high word. (There is no need for the mov_hi macro
to be separate)

0:     movw    %R0, #low offset >> 16
       adds    %Q0, %1, %R0, lsl #16
1:     mov     %R0, #high offset
       adc     %R0, %R0, #0
       .pushsection .pv_table,"a"
       .long 0b, 1b
       .popsection

The only problem is distinguishing the two mov instructions from each
other, but that should not be too hard I think.
Leizhen (ThunderTown) Sept. 21, 2020, 3:34 a.m. UTC | #6
On 2020/9/17 22:00, Ard Biesheuvel wrote:
> On Tue, 15 Sep 2020 at 22:06, Russell King - ARM Linux admin
> <linux@armlinux.org.uk> wrote:
>>
>> On Tue, Sep 15, 2020 at 09:16:15PM +0800, Zhen Lei wrote:
>>> Currently, only support the kernels where the base of physical memory is
>>> at a 16MiB boundary. Because the add/sub instructions only contains 8bits
>>> unrotated value. But we can use one more "add/sub" instructions to handle
>>> bits 23-16. The performance will be slightly affected.
>>>
>>> Since most boards meet 16 MiB alignment, so add a new configuration
>>> option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if
>>> anyone really needs it.
>>>
>>> All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are
>>> used in __fixup_a_pv_table() now, but the callee saved r11 is not used in
>>> the whole head.S file. So choose it.
>>>
>>> Because the calculation of "y = x + __pv_offset[63:24]" have been done,
>>> so we only need to calculate "y = y + __pv_offset[23:16]", that's why
>>> the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub()
>>> in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t"
>>> (above y).
>>>
>>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>>> ---
>>>  arch/arm/Kconfig              | 18 +++++++++++++++++-
>>>  arch/arm/include/asm/memory.h | 16 +++++++++++++---
>>>  arch/arm/kernel/head.S        | 25 +++++++++++++++++++------
>>>  3 files changed, 49 insertions(+), 10 deletions(-)
>>>
>>> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
>>> index e00d94b16658765..19fc2c746e2ce29 100644
>>> --- a/arch/arm/Kconfig
>>> +++ b/arch/arm/Kconfig
>>> @@ -240,12 +240,28 @@ config ARM_PATCH_PHYS_VIRT
>>>         kernel in system memory.
>>>
>>>         This can only be used with non-XIP MMU kernels where the base
>>> -       of physical memory is at a 16MB boundary.
>>> +       of physical memory is at a 16MiB boundary.
>>>
>>>         Only disable this option if you know that you do not require
>>>         this feature (eg, building a kernel for a single machine) and
>>>         you need to shrink the kernel to the minimal size.
>>>
>>> +config ARM_PATCH_PHYS_VIRT_RADICAL
>>> +     bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary"
>>> +     default n
>>
>> Please drop the "default n" - this is the default anyway.
>>
>>> @@ -236,6 +243,9 @@ static inline unsigned long __phys_to_virt(phys_addr_t x)
>>>        * in place where 'r' 32 bit operand is expected.
>>>        */
>>>       __pv_stub((unsigned long) x, t, "sub", __PV_BITS_31_24);
>>> +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
>>> +     __pv_stub((unsigned long) t, t, "sub", __PV_BITS_23_16);
>>
>> t is already unsigned long, so this cast is not necessary.
>>
>> I've been debating whether it would be better to use "movw" for this
>> for ARMv7.  In other words:
>>
>>         movw    tmp, #16-bit
>>         adds    %Q0, %1, tmp, lsl #16
>>         adc     %R0, %R0, #0
>>
>> It would certainly be less instructions, but at the cost of an
>> additional register - and we'd have to change the fixup code to
>> know about movw.
>>
>> Thoughts?
>>
> 
> Since LPAE implies v7, we can use movw unconditionally, which is nice.
> 
> There is no need to use an additional temp register, as we can use the
> register holding the high word. (There is no need for the mov_hi macro
> to be separate)
> 
> 0:     movw    %R0, #low offset >> 16
>        adds    %Q0, %1, %R0, lsl #16
> 1:     mov     %R0, #high offset
>        adc     %R0, %R0, #0
>        .pushsection .pv_table,"a"
>        .long 0b, 1b
>        .popsection
> 
> The only problem is distinguishing the two mov instructions from each

The #high offset can also consider use movw, it just save two bytes in
the thumb2 scenario. We can store different imm16 value for high_offset
and low_offset, so that we can distinguish them in __fixup_a_pv_table().

This will make the final implementation of the code look more clear and
consistent, especially THUMB2.

Let me try it.

> other, but that should not be too hard I think.
> 
> .
>
Ard Biesheuvel Sept. 21, 2020, 6:47 a.m. UTC | #7
On Mon, 21 Sep 2020 at 05:35, Leizhen (ThunderTown)
<thunder.leizhen@huawei.com> wrote:
>
>
>
> On 2020/9/17 22:00, Ard Biesheuvel wrote:
> > On Tue, 15 Sep 2020 at 22:06, Russell King - ARM Linux admin
> > <linux@armlinux.org.uk> wrote:
> >>
> >> On Tue, Sep 15, 2020 at 09:16:15PM +0800, Zhen Lei wrote:
> >>> Currently, only support the kernels where the base of physical memory is
> >>> at a 16MiB boundary. Because the add/sub instructions only contains 8bits
> >>> unrotated value. But we can use one more "add/sub" instructions to handle
> >>> bits 23-16. The performance will be slightly affected.
> >>>
> >>> Since most boards meet 16 MiB alignment, so add a new configuration
> >>> option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if
> >>> anyone really needs it.
> >>>
> >>> All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are
> >>> used in __fixup_a_pv_table() now, but the callee saved r11 is not used in
> >>> the whole head.S file. So choose it.
> >>>
> >>> Because the calculation of "y = x + __pv_offset[63:24]" have been done,
> >>> so we only need to calculate "y = y + __pv_offset[23:16]", that's why
> >>> the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub()
> >>> in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t"
> >>> (above y).
> >>>
> >>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
> >>> ---
> >>>  arch/arm/Kconfig              | 18 +++++++++++++++++-
> >>>  arch/arm/include/asm/memory.h | 16 +++++++++++++---
> >>>  arch/arm/kernel/head.S        | 25 +++++++++++++++++++------
> >>>  3 files changed, 49 insertions(+), 10 deletions(-)
> >>>
> >>> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
> >>> index e00d94b16658765..19fc2c746e2ce29 100644
> >>> --- a/arch/arm/Kconfig
> >>> +++ b/arch/arm/Kconfig
> >>> @@ -240,12 +240,28 @@ config ARM_PATCH_PHYS_VIRT
> >>>         kernel in system memory.
> >>>
> >>>         This can only be used with non-XIP MMU kernels where the base
> >>> -       of physical memory is at a 16MB boundary.
> >>> +       of physical memory is at a 16MiB boundary.
> >>>
> >>>         Only disable this option if you know that you do not require
> >>>         this feature (eg, building a kernel for a single machine) and
> >>>         you need to shrink the kernel to the minimal size.
> >>>
> >>> +config ARM_PATCH_PHYS_VIRT_RADICAL
> >>> +     bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary"
> >>> +     default n
> >>
> >> Please drop the "default n" - this is the default anyway.
> >>
> >>> @@ -236,6 +243,9 @@ static inline unsigned long __phys_to_virt(phys_addr_t x)
> >>>        * in place where 'r' 32 bit operand is expected.
> >>>        */
> >>>       __pv_stub((unsigned long) x, t, "sub", __PV_BITS_31_24);
> >>> +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
> >>> +     __pv_stub((unsigned long) t, t, "sub", __PV_BITS_23_16);
> >>
> >> t is already unsigned long, so this cast is not necessary.
> >>
> >> I've been debating whether it would be better to use "movw" for this
> >> for ARMv7.  In other words:
> >>
> >>         movw    tmp, #16-bit
> >>         adds    %Q0, %1, tmp, lsl #16
> >>         adc     %R0, %R0, #0
> >>
> >> It would certainly be less instructions, but at the cost of an
> >> additional register - and we'd have to change the fixup code to
> >> know about movw.
> >>
> >> Thoughts?
> >>
> >
> > Since LPAE implies v7, we can use movw unconditionally, which is nice.
> >
> > There is no need to use an additional temp register, as we can use the
> > register holding the high word. (There is no need for the mov_hi macro
> > to be separate)
> >
> > 0:     movw    %R0, #low offset >> 16
> >        adds    %Q0, %1, %R0, lsl #16
> > 1:     mov     %R0, #high offset
> >        adc     %R0, %R0, #0
> >        .pushsection .pv_table,"a"
> >        .long 0b, 1b
> >        .popsection
> >
> > The only problem is distinguishing the two mov instructions from each
>
> The #high offset can also consider use movw, it just save two bytes in
> the thumb2 scenario. We can store different imm16 value for high_offset
> and low_offset, so that we can distinguish them in __fixup_a_pv_table().
>
> This will make the final implementation of the code look more clear and
> consistent, especially THUMB2.
>
> Let me try it.
>

Hello Zhen Lei,

I am looking into this as well:

https://git.kernel.org/pub/scm/linux/kernel/git/ardb/linux.git/log/?h=arm-p2v-v2

Could you please test this version on your hardware?
Leizhen (ThunderTown) Sept. 21, 2020, 8:53 a.m. UTC | #8
On 2020/9/21 14:47, Ard Biesheuvel wrote:
> On Mon, 21 Sep 2020 at 05:35, Leizhen (ThunderTown)
> <thunder.leizhen@huawei.com> wrote:
>>
>>
>>
>> On 2020/9/17 22:00, Ard Biesheuvel wrote:
>>> On Tue, 15 Sep 2020 at 22:06, Russell King - ARM Linux admin
>>> <linux@armlinux.org.uk> wrote:
>>>>
>>>> On Tue, Sep 15, 2020 at 09:16:15PM +0800, Zhen Lei wrote:
>>>>> Currently, only support the kernels where the base of physical memory is
>>>>> at a 16MiB boundary. Because the add/sub instructions only contains 8bits
>>>>> unrotated value. But we can use one more "add/sub" instructions to handle
>>>>> bits 23-16. The performance will be slightly affected.
>>>>>
>>>>> Since most boards meet 16 MiB alignment, so add a new configuration
>>>>> option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if
>>>>> anyone really needs it.
>>>>>
>>>>> All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are
>>>>> used in __fixup_a_pv_table() now, but the callee saved r11 is not used in
>>>>> the whole head.S file. So choose it.
>>>>>
>>>>> Because the calculation of "y = x + __pv_offset[63:24]" have been done,
>>>>> so we only need to calculate "y = y + __pv_offset[23:16]", that's why
>>>>> the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub()
>>>>> in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t"
>>>>> (above y).
>>>>>
>>>>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>>>>> ---
>>>>>  arch/arm/Kconfig              | 18 +++++++++++++++++-
>>>>>  arch/arm/include/asm/memory.h | 16 +++++++++++++---
>>>>>  arch/arm/kernel/head.S        | 25 +++++++++++++++++++------
>>>>>  3 files changed, 49 insertions(+), 10 deletions(-)
>>>>>
>>>>> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
>>>>> index e00d94b16658765..19fc2c746e2ce29 100644
>>>>> --- a/arch/arm/Kconfig
>>>>> +++ b/arch/arm/Kconfig
>>>>> @@ -240,12 +240,28 @@ config ARM_PATCH_PHYS_VIRT
>>>>>         kernel in system memory.
>>>>>
>>>>>         This can only be used with non-XIP MMU kernels where the base
>>>>> -       of physical memory is at a 16MB boundary.
>>>>> +       of physical memory is at a 16MiB boundary.
>>>>>
>>>>>         Only disable this option if you know that you do not require
>>>>>         this feature (eg, building a kernel for a single machine) and
>>>>>         you need to shrink the kernel to the minimal size.
>>>>>
>>>>> +config ARM_PATCH_PHYS_VIRT_RADICAL
>>>>> +     bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary"
>>>>> +     default n
>>>>
>>>> Please drop the "default n" - this is the default anyway.
>>>>
>>>>> @@ -236,6 +243,9 @@ static inline unsigned long __phys_to_virt(phys_addr_t x)
>>>>>        * in place where 'r' 32 bit operand is expected.
>>>>>        */
>>>>>       __pv_stub((unsigned long) x, t, "sub", __PV_BITS_31_24);
>>>>> +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
>>>>> +     __pv_stub((unsigned long) t, t, "sub", __PV_BITS_23_16);
>>>>
>>>> t is already unsigned long, so this cast is not necessary.
>>>>
>>>> I've been debating whether it would be better to use "movw" for this
>>>> for ARMv7.  In other words:
>>>>
>>>>         movw    tmp, #16-bit
>>>>         adds    %Q0, %1, tmp, lsl #16
>>>>         adc     %R0, %R0, #0
>>>>
>>>> It would certainly be less instructions, but at the cost of an
>>>> additional register - and we'd have to change the fixup code to
>>>> know about movw.
>>>>
>>>> Thoughts?
>>>>
>>>
>>> Since LPAE implies v7, we can use movw unconditionally, which is nice.
>>>
>>> There is no need to use an additional temp register, as we can use the
>>> register holding the high word. (There is no need for the mov_hi macro
>>> to be separate)
>>>
>>> 0:     movw    %R0, #low offset >> 16
>>>        adds    %Q0, %1, %R0, lsl #16
>>> 1:     mov     %R0, #high offset
>>>        adc     %R0, %R0, #0
>>>        .pushsection .pv_table,"a"
>>>        .long 0b, 1b
>>>        .popsection
>>>
>>> The only problem is distinguishing the two mov instructions from each
>>
>> The #high offset can also consider use movw, it just save two bytes in
>> the thumb2 scenario. We can store different imm16 value for high_offset
>> and low_offset, so that we can distinguish them in __fixup_a_pv_table().
>>
>> This will make the final implementation of the code look more clear and
>> consistent, especially THUMB2.
>>
>> Let me try it.
>>
> 
> Hello Zhen Lei,
> 
> I am looking into this as well:
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/ardb/linux.git/log/?h=arm-p2v-v2
> 
> Could you please test this version on your hardware?

OK, I will test it on my boards.

> 
> .
>
Leizhen (ThunderTown) Sept. 22, 2020, 12:30 p.m. UTC | #9
On 2020/9/21 16:53, Leizhen (ThunderTown) wrote:
> 
> 
> On 2020/9/21 14:47, Ard Biesheuvel wrote:
>> On Mon, 21 Sep 2020 at 05:35, Leizhen (ThunderTown)
>> <thunder.leizhen@huawei.com> wrote:
>>>
>>>
>>>
>>> On 2020/9/17 22:00, Ard Biesheuvel wrote:
>>>> On Tue, 15 Sep 2020 at 22:06, Russell King - ARM Linux admin
>>>> <linux@armlinux.org.uk> wrote:
>>>>>
>>>>> On Tue, Sep 15, 2020 at 09:16:15PM +0800, Zhen Lei wrote:
>>>>>> Currently, only support the kernels where the base of physical memory is
>>>>>> at a 16MiB boundary. Because the add/sub instructions only contains 8bits
>>>>>> unrotated value. But we can use one more "add/sub" instructions to handle
>>>>>> bits 23-16. The performance will be slightly affected.
>>>>>>
>>>>>> Since most boards meet 16 MiB alignment, so add a new configuration
>>>>>> option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if
>>>>>> anyone really needs it.
>>>>>>
>>>>>> All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are
>>>>>> used in __fixup_a_pv_table() now, but the callee saved r11 is not used in
>>>>>> the whole head.S file. So choose it.
>>>>>>
>>>>>> Because the calculation of "y = x + __pv_offset[63:24]" have been done,
>>>>>> so we only need to calculate "y = y + __pv_offset[23:16]", that's why
>>>>>> the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub()
>>>>>> in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t"
>>>>>> (above y).
>>>>>>
>>>>>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>>>>>> ---
>>>>>>  arch/arm/Kconfig              | 18 +++++++++++++++++-
>>>>>>  arch/arm/include/asm/memory.h | 16 +++++++++++++---
>>>>>>  arch/arm/kernel/head.S        | 25 +++++++++++++++++++------
>>>>>>  3 files changed, 49 insertions(+), 10 deletions(-)
>>>>>>
>>>>>> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
>>>>>> index e00d94b16658765..19fc2c746e2ce29 100644
>>>>>> --- a/arch/arm/Kconfig
>>>>>> +++ b/arch/arm/Kconfig
>>>>>> @@ -240,12 +240,28 @@ config ARM_PATCH_PHYS_VIRT
>>>>>>         kernel in system memory.
>>>>>>
>>>>>>         This can only be used with non-XIP MMU kernels where the base
>>>>>> -       of physical memory is at a 16MB boundary.
>>>>>> +       of physical memory is at a 16MiB boundary.
>>>>>>
>>>>>>         Only disable this option if you know that you do not require
>>>>>>         this feature (eg, building a kernel for a single machine) and
>>>>>>         you need to shrink the kernel to the minimal size.
>>>>>>
>>>>>> +config ARM_PATCH_PHYS_VIRT_RADICAL
>>>>>> +     bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary"
>>>>>> +     default n
>>>>>
>>>>> Please drop the "default n" - this is the default anyway.
>>>>>
>>>>>> @@ -236,6 +243,9 @@ static inline unsigned long __phys_to_virt(phys_addr_t x)
>>>>>>        * in place where 'r' 32 bit operand is expected.
>>>>>>        */
>>>>>>       __pv_stub((unsigned long) x, t, "sub", __PV_BITS_31_24);
>>>>>> +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
>>>>>> +     __pv_stub((unsigned long) t, t, "sub", __PV_BITS_23_16);
>>>>>
>>>>> t is already unsigned long, so this cast is not necessary.
>>>>>
>>>>> I've been debating whether it would be better to use "movw" for this
>>>>> for ARMv7.  In other words:
>>>>>
>>>>>         movw    tmp, #16-bit
>>>>>         adds    %Q0, %1, tmp, lsl #16
>>>>>         adc     %R0, %R0, #0
>>>>>
>>>>> It would certainly be less instructions, but at the cost of an
>>>>> additional register - and we'd have to change the fixup code to
>>>>> know about movw.
>>>>>
>>>>> Thoughts?
>>>>>
>>>>
>>>> Since LPAE implies v7, we can use movw unconditionally, which is nice.
>>>>
>>>> There is no need to use an additional temp register, as we can use the
>>>> register holding the high word. (There is no need for the mov_hi macro
>>>> to be separate)
>>>>
>>>> 0:     movw    %R0, #low offset >> 16
>>>>        adds    %Q0, %1, %R0, lsl #16
>>>> 1:     mov     %R0, #high offset
>>>>        adc     %R0, %R0, #0
>>>>        .pushsection .pv_table,"a"
>>>>        .long 0b, 1b
>>>>        .popsection
>>>>
>>>> The only problem is distinguishing the two mov instructions from each
>>>
>>> The #high offset can also consider use movw, it just save two bytes in
>>> the thumb2 scenario. We can store different imm16 value for high_offset
>>> and low_offset, so that we can distinguish them in __fixup_a_pv_table().
>>>
>>> This will make the final implementation of the code look more clear and
>>> consistent, especially THUMB2.
>>>
>>> Let me try it.
>>>
>>
>> Hello Zhen Lei,
>>
>> I am looking into this as well:
>>
>> https://git.kernel.org/pub/scm/linux/kernel/git/ardb/linux.git/log/?h=arm-p2v-v2
>>
>> Could you please test this version on your hardware?
> 
> OK, I will test it on my boards.
Hi Ard Biesheuvel:
  I have tested it on 16MiB aligned + LE board, it works well. I've asked my colleagues
from other departments to run it on 2MiB aligned + BE board. He will do it tomorrow.


> 
>>
>> .
>>
> 
> 
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
> 
> .
>
Leizhen (ThunderTown) Sept. 28, 2020, 1:30 a.m. UTC | #10
On 2020/9/22 20:30, Leizhen (ThunderTown) wrote:
> 
> 
> On 2020/9/21 16:53, Leizhen (ThunderTown) wrote:
>>
>>
>> On 2020/9/21 14:47, Ard Biesheuvel wrote:
>>> On Mon, 21 Sep 2020 at 05:35, Leizhen (ThunderTown)
>>> <thunder.leizhen@huawei.com> wrote:
>>>>
>>>>
>>>>
>>>> On 2020/9/17 22:00, Ard Biesheuvel wrote:
>>>>> On Tue, 15 Sep 2020 at 22:06, Russell King - ARM Linux admin
>>>>> <linux@armlinux.org.uk> wrote:
>>>>>>
>>>>>> On Tue, Sep 15, 2020 at 09:16:15PM +0800, Zhen Lei wrote:
>>>>>>> Currently, only support the kernels where the base of physical memory is
>>>>>>> at a 16MiB boundary. Because the add/sub instructions only contains 8bits
>>>>>>> unrotated value. But we can use one more "add/sub" instructions to handle
>>>>>>> bits 23-16. The performance will be slightly affected.
>>>>>>>
>>>>>>> Since most boards meet 16 MiB alignment, so add a new configuration
>>>>>>> option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if
>>>>>>> anyone really needs it.
>>>>>>>
>>>>>>> All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are
>>>>>>> used in __fixup_a_pv_table() now, but the callee saved r11 is not used in
>>>>>>> the whole head.S file. So choose it.
>>>>>>>
>>>>>>> Because the calculation of "y = x + __pv_offset[63:24]" have been done,
>>>>>>> so we only need to calculate "y = y + __pv_offset[23:16]", that's why
>>>>>>> the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub()
>>>>>>> in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t"
>>>>>>> (above y).
>>>>>>>
>>>>>>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>>>>>>> ---
>>>>>>>  arch/arm/Kconfig              | 18 +++++++++++++++++-
>>>>>>>  arch/arm/include/asm/memory.h | 16 +++++++++++++---
>>>>>>>  arch/arm/kernel/head.S        | 25 +++++++++++++++++++------
>>>>>>>  3 files changed, 49 insertions(+), 10 deletions(-)
>>>>>>>
>>>>>>> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
>>>>>>> index e00d94b16658765..19fc2c746e2ce29 100644
>>>>>>> --- a/arch/arm/Kconfig
>>>>>>> +++ b/arch/arm/Kconfig
>>>>>>> @@ -240,12 +240,28 @@ config ARM_PATCH_PHYS_VIRT
>>>>>>>         kernel in system memory.
>>>>>>>
>>>>>>>         This can only be used with non-XIP MMU kernels where the base
>>>>>>> -       of physical memory is at a 16MB boundary.
>>>>>>> +       of physical memory is at a 16MiB boundary.
>>>>>>>
>>>>>>>         Only disable this option if you know that you do not require
>>>>>>>         this feature (eg, building a kernel for a single machine) and
>>>>>>>         you need to shrink the kernel to the minimal size.
>>>>>>>
>>>>>>> +config ARM_PATCH_PHYS_VIRT_RADICAL
>>>>>>> +     bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary"
>>>>>>> +     default n
>>>>>>
>>>>>> Please drop the "default n" - this is the default anyway.
>>>>>>
>>>>>>> @@ -236,6 +243,9 @@ static inline unsigned long __phys_to_virt(phys_addr_t x)
>>>>>>>        * in place where 'r' 32 bit operand is expected.
>>>>>>>        */
>>>>>>>       __pv_stub((unsigned long) x, t, "sub", __PV_BITS_31_24);
>>>>>>> +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
>>>>>>> +     __pv_stub((unsigned long) t, t, "sub", __PV_BITS_23_16);
>>>>>>
>>>>>> t is already unsigned long, so this cast is not necessary.
>>>>>>
>>>>>> I've been debating whether it would be better to use "movw" for this
>>>>>> for ARMv7.  In other words:
>>>>>>
>>>>>>         movw    tmp, #16-bit
>>>>>>         adds    %Q0, %1, tmp, lsl #16
>>>>>>         adc     %R0, %R0, #0
>>>>>>
>>>>>> It would certainly be less instructions, but at the cost of an
>>>>>> additional register - and we'd have to change the fixup code to
>>>>>> know about movw.
>>>>>>
>>>>>> Thoughts?
>>>>>>
>>>>>
>>>>> Since LPAE implies v7, we can use movw unconditionally, which is nice.
>>>>>
>>>>> There is no need to use an additional temp register, as we can use the
>>>>> register holding the high word. (There is no need for the mov_hi macro
>>>>> to be separate)
>>>>>
>>>>> 0:     movw    %R0, #low offset >> 16
>>>>>        adds    %Q0, %1, %R0, lsl #16
>>>>> 1:     mov     %R0, #high offset
>>>>>        adc     %R0, %R0, #0
>>>>>        .pushsection .pv_table,"a"
>>>>>        .long 0b, 1b
>>>>>        .popsection
>>>>>
>>>>> The only problem is distinguishing the two mov instructions from each
>>>>
>>>> The #high offset can also consider use movw, it just save two bytes in
>>>> the thumb2 scenario. We can store different imm16 value for high_offset
>>>> and low_offset, so that we can distinguish them in __fixup_a_pv_table().
>>>>
>>>> This will make the final implementation of the code look more clear and
>>>> consistent, especially THUMB2.
>>>>
>>>> Let me try it.
>>>>
>>>
>>> Hello Zhen Lei,
>>>
>>> I am looking into this as well:
>>>
>>> https://git.kernel.org/pub/scm/linux/kernel/git/ardb/linux.git/log/?h=arm-p2v-v2
>>>
>>> Could you please test this version on your hardware?
>>
>> OK, I will test it on my boards.
> Hi Ard Biesheuvel:
>   I have tested it on 16MiB aligned + LE board, it works well. I've asked my colleagues
> from other departments to run it on 2MiB aligned + BE board. He will do it tomorrow.

Hi, Ard Biesheuvel:
  I'm sorry to keep you waiting so long. You patch series works well on 2MiB aligned + BE board
also. I spent a lot of time, because our 2MiB aligned + BE board loads zImage. Therefore, special
processing is required for the following code:

arch/arm/boot/compressed/head.S:
#ifdef CONFIG_AUTO_ZRELADDR
                mov     r4, pc
                and     r4, r4, #0xf8000000		//currently only support 128MiB alignment
                add     r4, r4, #TEXT_OFFSET
#else

This is a special scenario that does not conflict with your code framework. So I'm trying to fix it.

Tested-by: Zhen Lei <thunder.leizhen@huawei.com>


> 
> 
>>
>>>
>>> .
>>>
>>
>>
>> _______________________________________________
>> linux-arm-kernel mailing list
>> linux-arm-kernel@lists.infradead.org
>> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
>>
>> .
>>
Leizhen (ThunderTown) Sept. 28, 2020, 9:30 a.m. UTC | #11
On 2020/9/28 9:30, Leizhen (ThunderTown) wrote:
> 
> 
> On 2020/9/22 20:30, Leizhen (ThunderTown) wrote:
>>
>>
>> On 2020/9/21 16:53, Leizhen (ThunderTown) wrote:
>>>
>>>
>>> On 2020/9/21 14:47, Ard Biesheuvel wrote:
>>>> On Mon, 21 Sep 2020 at 05:35, Leizhen (ThunderTown)
>>>> <thunder.leizhen@huawei.com> wrote:
>>>>>
>>>>>
>>>>>
>>>>> On 2020/9/17 22:00, Ard Biesheuvel wrote:
>>>>>> On Tue, 15 Sep 2020 at 22:06, Russell King - ARM Linux admin
>>>>>> <linux@armlinux.org.uk> wrote:
>>>>>>>
>>>>>>> On Tue, Sep 15, 2020 at 09:16:15PM +0800, Zhen Lei wrote:
>>>>>>>> Currently, only support the kernels where the base of physical memory is
>>>>>>>> at a 16MiB boundary. Because the add/sub instructions only contains 8bits
>>>>>>>> unrotated value. But we can use one more "add/sub" instructions to handle
>>>>>>>> bits 23-16. The performance will be slightly affected.
>>>>>>>>
>>>>>>>> Since most boards meet 16 MiB alignment, so add a new configuration
>>>>>>>> option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if
>>>>>>>> anyone really needs it.
>>>>>>>>
>>>>>>>> All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are
>>>>>>>> used in __fixup_a_pv_table() now, but the callee saved r11 is not used in
>>>>>>>> the whole head.S file. So choose it.
>>>>>>>>
>>>>>>>> Because the calculation of "y = x + __pv_offset[63:24]" have been done,
>>>>>>>> so we only need to calculate "y = y + __pv_offset[23:16]", that's why
>>>>>>>> the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub()
>>>>>>>> in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t"
>>>>>>>> (above y).
>>>>>>>>
>>>>>>>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
>>>>>>>> ---
>>>>>>>>  arch/arm/Kconfig              | 18 +++++++++++++++++-
>>>>>>>>  arch/arm/include/asm/memory.h | 16 +++++++++++++---
>>>>>>>>  arch/arm/kernel/head.S        | 25 +++++++++++++++++++------
>>>>>>>>  3 files changed, 49 insertions(+), 10 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
>>>>>>>> index e00d94b16658765..19fc2c746e2ce29 100644
>>>>>>>> --- a/arch/arm/Kconfig
>>>>>>>> +++ b/arch/arm/Kconfig
>>>>>>>> @@ -240,12 +240,28 @@ config ARM_PATCH_PHYS_VIRT
>>>>>>>>         kernel in system memory.
>>>>>>>>
>>>>>>>>         This can only be used with non-XIP MMU kernels where the base
>>>>>>>> -       of physical memory is at a 16MB boundary.
>>>>>>>> +       of physical memory is at a 16MiB boundary.
>>>>>>>>
>>>>>>>>         Only disable this option if you know that you do not require
>>>>>>>>         this feature (eg, building a kernel for a single machine) and
>>>>>>>>         you need to shrink the kernel to the minimal size.
>>>>>>>>
>>>>>>>> +config ARM_PATCH_PHYS_VIRT_RADICAL
>>>>>>>> +     bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary"
>>>>>>>> +     default n
>>>>>>>
>>>>>>> Please drop the "default n" - this is the default anyway.
>>>>>>>
>>>>>>>> @@ -236,6 +243,9 @@ static inline unsigned long __phys_to_virt(phys_addr_t x)
>>>>>>>>        * in place where 'r' 32 bit operand is expected.
>>>>>>>>        */
>>>>>>>>       __pv_stub((unsigned long) x, t, "sub", __PV_BITS_31_24);
>>>>>>>> +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
>>>>>>>> +     __pv_stub((unsigned long) t, t, "sub", __PV_BITS_23_16);
>>>>>>>
>>>>>>> t is already unsigned long, so this cast is not necessary.
>>>>>>>
>>>>>>> I've been debating whether it would be better to use "movw" for this
>>>>>>> for ARMv7.  In other words:
>>>>>>>
>>>>>>>         movw    tmp, #16-bit
>>>>>>>         adds    %Q0, %1, tmp, lsl #16
>>>>>>>         adc     %R0, %R0, #0
>>>>>>>
>>>>>>> It would certainly be less instructions, but at the cost of an
>>>>>>> additional register - and we'd have to change the fixup code to
>>>>>>> know about movw.
>>>>>>>
>>>>>>> Thoughts?
>>>>>>>
>>>>>>
>>>>>> Since LPAE implies v7, we can use movw unconditionally, which is nice.
>>>>>>
>>>>>> There is no need to use an additional temp register, as we can use the
>>>>>> register holding the high word. (There is no need for the mov_hi macro
>>>>>> to be separate)
>>>>>>
>>>>>> 0:     movw    %R0, #low offset >> 16
>>>>>>        adds    %Q0, %1, %R0, lsl #16
>>>>>> 1:     mov     %R0, #high offset
>>>>>>        adc     %R0, %R0, #0
>>>>>>        .pushsection .pv_table,"a"
>>>>>>        .long 0b, 1b
>>>>>>        .popsection
>>>>>>
>>>>>> The only problem is distinguishing the two mov instructions from each
>>>>>
>>>>> The #high offset can also consider use movw, it just save two bytes in
>>>>> the thumb2 scenario. We can store different imm16 value for high_offset
>>>>> and low_offset, so that we can distinguish them in __fixup_a_pv_table().
>>>>>
>>>>> This will make the final implementation of the code look more clear and
>>>>> consistent, especially THUMB2.
>>>>>
>>>>> Let me try it.
>>>>>
>>>>
>>>> Hello Zhen Lei,
>>>>
>>>> I am looking into this as well:
>>>>
>>>> https://git.kernel.org/pub/scm/linux/kernel/git/ardb/linux.git/log/?h=arm-p2v-v2
>>>>
>>>> Could you please test this version on your hardware?
>>>
>>> OK, I will test it on my boards.
>> Hi Ard Biesheuvel:
>>   I have tested it on 16MiB aligned + LE board, it works well. I've asked my colleagues
>> from other departments to run it on 2MiB aligned + BE board. He will do it tomorrow.
> 
> Hi, Ard Biesheuvel:
>   I'm sorry to keep you waiting so long. You patch series works well on 2MiB aligned + BE board
> also. I spent a lot of time, because our 2MiB aligned + BE board loads zImage. Therefore, special
> processing is required for the following code:
> 
> arch/arm/boot/compressed/head.S:
> #ifdef CONFIG_AUTO_ZRELADDR
>                 mov     r4, pc
>                 and     r4, r4, #0xf8000000		//currently only support 128MiB alignment
>                 add     r4, r4, #TEXT_OFFSET
> #else
> 
> This is a special scenario that does not conflict with your code framework. So I'm trying to fix it.
> 
> Tested-by: Zhen Lei <thunder.leizhen@huawei.com>

Hi, Ard Biesheuvel:
  I just sent the above problem's fix patch.

  [PATCH 0/2] ARM: decompressor: relax the loading restriction of the decompressed kernel

> 
> 
>>
>>
>>>
>>>>
>>>> .
>>>>
>>>
>>>
>>> _______________________________________________
>>> linux-arm-kernel mailing list
>>> linux-arm-kernel@lists.infradead.org
>>> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
>>>
>>> .
>>>
diff mbox series

Patch

diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
index e00d94b16658765..19fc2c746e2ce29 100644
--- a/arch/arm/Kconfig
+++ b/arch/arm/Kconfig
@@ -240,12 +240,28 @@  config ARM_PATCH_PHYS_VIRT
 	  kernel in system memory.
 
 	  This can only be used with non-XIP MMU kernels where the base
-	  of physical memory is at a 16MB boundary.
+	  of physical memory is at a 16MiB boundary.
 
 	  Only disable this option if you know that you do not require
 	  this feature (eg, building a kernel for a single machine) and
 	  you need to shrink the kernel to the minimal size.
 
+config ARM_PATCH_PHYS_VIRT_RADICAL
+	bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary"
+	default n
+	depends on ARM_PATCH_PHYS_VIRT
+	depends on !THUMB2_KERNEL
+	help
+	  This can only be used with non-XIP MMU kernels where the base
+	  of physical memory is at a 64KiB boundary.
+
+	  Compared with ARM_PATCH_PHYS_VIRT, one or two more instructions
+	  need to be added to implement the conversion of bits 23-16 of
+	  the VA/PA in phys-to-virt and virt-to-phys. The performance is
+	  slightly affected.
+
+	  If unsure say N here.
+
 config NEED_MACH_IO_H
 	bool
 	help
diff --git a/arch/arm/include/asm/memory.h b/arch/arm/include/asm/memory.h
index 99035b5891ef442..71b3a60eeb1b1c6 100644
--- a/arch/arm/include/asm/memory.h
+++ b/arch/arm/include/asm/memory.h
@@ -173,6 +173,7 @@ 
  * so that all we need to do is modify the 8-bit constant field.
  */
 #define __PV_BITS_31_24	0x81000000
+#define __PV_BITS_23_16	0x00810000
 #define __PV_BITS_7_0	0x81
 
 extern unsigned long __pv_phys_pfn_offset;
@@ -201,7 +202,7 @@ 
 	: "=r" (t)					\
 	: "I" (__PV_BITS_7_0))
 
-#define __pv_add_carry_stub(x, y)			\
+#define __pv_add_carry_stub(x, y, type)			\
 	__asm__ volatile("@ __pv_add_carry_stub\n"	\
 	"1:	adds	%Q0, %1, %2\n"			\
 	"	adc	%R0, %R0, #0\n"			\
@@ -209,7 +210,7 @@ 
 	"	.long	1b\n"				\
 	"	.popsection\n"				\
 	: "+r" (y)					\
-	: "r" (x), "I" (__PV_BITS_31_24)		\
+	: "r" (x), "I" (type)				\
 	: "cc")
 
 static inline phys_addr_t __virt_to_phys_nodebug(unsigned long x)
@@ -218,9 +219,15 @@  static inline phys_addr_t __virt_to_phys_nodebug(unsigned long x)
 
 	if (sizeof(phys_addr_t) == 4) {
 		__pv_stub(x, t, "add", __PV_BITS_31_24);
+#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
+		__pv_stub(t, t, "add", __PV_BITS_23_16);
+#endif
 	} else {
 		__pv_stub_mov_hi(t);
-		__pv_add_carry_stub(x, t);
+		__pv_add_carry_stub(x, t, __PV_BITS_31_24);
+#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
+		__pv_add_carry_stub(t, t, __PV_BITS_23_16);
+#endif
 	}
 	return t;
 }
@@ -236,6 +243,9 @@  static inline unsigned long __phys_to_virt(phys_addr_t x)
 	 * in place where 'r' 32 bit operand is expected.
 	 */
 	__pv_stub((unsigned long) x, t, "sub", __PV_BITS_31_24);
+#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
+	__pv_stub((unsigned long) t, t, "sub", __PV_BITS_23_16);
+#endif
 	return t;
 }
 
diff --git a/arch/arm/kernel/head.S b/arch/arm/kernel/head.S
index 02d78c9198d0e8d..d9fb226a24d43ae 100644
--- a/arch/arm/kernel/head.S
+++ b/arch/arm/kernel/head.S
@@ -120,7 +120,7 @@  ENTRY(stext)
 	bl	__fixup_smp
 #endif
 #ifdef CONFIG_ARM_PATCH_PHYS_VIRT
-	bl	__fixup_pv_table
+	bl	__fixup_pv_table		@r11 will be used
 #endif
 	bl	__create_page_tables
 
@@ -614,8 +614,13 @@  __fixup_pv_table:
 	mov	r0, r8, lsr #PAGE_SHIFT	@ convert to PFN
 	str	r0, [r6]	@ save computed PHYS_PFN_OFFSET to __pv_phys_pfn_offset
 	strcc	ip, [r7, #HIGH_OFFSET]	@ save to __pv_offset high bits
+#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL
+	mov	r6, r3, lsr #16	@ constant for add/sub instructions
+	teq	r3, r6, lsl #16 @ must be 64KiB aligned
+#else
 	mov	r6, r3, lsr #24	@ constant for add/sub instructions
 	teq	r3, r6, lsl #24 @ must be 16MiB aligned
+#endif
 THUMB(	it	ne		@ cross section branch )
 	bne	__error
 	str	r3, [r7, #LOW_OFFSET]	@ save to __pv_offset low bits
@@ -636,7 +641,9 @@  __fixup_a_pv_table:
 	add	r6, r6, r3
 	ldr	r0, [r6, #HIGH_OFFSET]	@ __pv_offset high word
 	ldr	r6, [r6, #LOW_OFFSET]	@ __pv_offset low word
-	mov	r6, r6, lsr #24
+	mov	r11, r6, lsl #8
+	mov	r11, r11, lsr #24	@ bits 23-16
+	mov	r6, r6, lsr #24		@ bits 31-24
 	cmn	r0, #1
 #ifdef CONFIG_THUMB2_KERNEL
 	moveq	r0, #0x200000	@ set bit 21, mov to mvn instruction
@@ -682,14 +689,20 @@  ARM_BE8(rev16	ip, ip)
 #ifdef CONFIG_CPU_ENDIAN_BE8
 	@ in BE8, we load data in BE, but instructions still in LE
 	bic	ip, ip, #0xff000000
-	tst	ip, #0x000f0000	@ check the rotation field
+	tst	ip, #0x00040000	@ check the rotation field
 	orrne	ip, ip, r6, lsl #24 @ mask in offset bits 31-24
+	tst	ip, #0x00080000	@ check the rotation field
+	orrne	ip, ip, r11, lsl #24 @ mask in offset bits 23-16
+	tst	ip, #0x000f0000	@ check the rotation field
 	biceq	ip, ip, #0x00004000 @ clear bit 22
 	orreq	ip, ip, r0      @ mask in offset bits 7-0
 #else
 	bic	ip, ip, #0x000000ff
-	tst	ip, #0xf00	@ check the rotation field
+	tst	ip, #0x400	@ check the rotation field
 	orrne	ip, ip, r6	@ mask in offset bits 31-24
+	tst	ip, #0x800	@ check the rotation field
+	orrne	ip, ip, r11	@ mask in offset bits 23-16
+	tst	ip, #0xf00	@ check the rotation field
 	biceq	ip, ip, #0x400000	@ clear bit 22
 	orreq	ip, ip, r0	@ mask in offset bits 7-0
 #endif
@@ -705,12 +718,12 @@  ENDPROC(__fixup_a_pv_table)
 3:	.long __pv_offset
 
 ENTRY(fixup_pv_table)
-	stmfd	sp!, {r4 - r7, lr}
+	stmfd	sp!, {r4 - r7, r11, lr}
 	mov	r3, #0			@ no offset
 	mov	r4, r0			@ r0 = table start
 	add	r5, r0, r1		@ r1 = table size
 	bl	__fixup_a_pv_table
-	ldmfd	sp!, {r4 - r7, pc}
+	ldmfd	sp!, {r4 - r7, r11, pc}
 ENDPROC(fixup_pv_table)
 
 	.data