[PATCH] arm64: add architecture specified current_pt_regs

[PATCH] arm64: add architecture specified current_pt_regs

[Zhi-zhou Zhang]

From: zhizhou <zhizhou.zh@gmail.com>

This patch is based on the implementation of arm. The generic
current_pt_regs is implemented with current->stack. It need to access
memory that would be too expensive.

Signed-off-by: zhizhou <zhizhou.zh@gmail.com>
---
 arch/arm64/include/asm/ptrace.h | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/arch/arm64/include/asm/ptrace.h b/arch/arm64/include/asm/ptrace.h
index e9e5467..1865d54 100644
--- a/arch/arm64/include/asm/ptrace.h
+++ b/arch/arm64/include/asm/ptrace.h
@@ -185,5 +185,9 @@ static inline int valid_user_regs(struct user_pt_regs *regs)
 
 extern unsigned long profile_pc(struct pt_regs *regs);
 
+#define current_pt_regs(void) ({ (struct pt_regs *)			\
+	((current_stack_pointer | (THREAD_SIZE - 1)) - 0xf) - 1;	\
+})
+
 #endif /* __ASSEMBLY__ */
 #endif
-- 
2.5.0

Re: [PATCH] arm64: add architecture specified current_pt_regs

[Catalin Marinas]

On Thu, Feb 18, 2016 at 07:48:35PM +0800, Zhi-zhou Zhang wrote:
> From: zhizhou <zhizhou.zh@gmail.com>
> 
> This patch is based on the implementation of arm. The generic
> current_pt_regs is implemented with current->stack. It need to access
> memory that would be too expensive.

Do you have any performance numbers?

> diff --git a/arch/arm64/include/asm/ptrace.h b/arch/arm64/include/asm/ptrace.h
> index e9e5467..1865d54 100644
> --- a/arch/arm64/include/asm/ptrace.h
> +++ b/arch/arm64/include/asm/ptrace.h
> @@ -185,5 +185,9 @@ static inline int valid_user_regs(struct user_pt_regs *regs)
>  
>  extern unsigned long profile_pc(struct pt_regs *regs);
>  
> +#define current_pt_regs(void) ({ (struct pt_regs *)			\
> +	((current_stack_pointer | (THREAD_SIZE - 1)) - 0xf) - 1;	\
> +})

I don't think this works well with the separate IRQ stack that we merged
in 4.5-rc1. current_thread_info() explicitly uses "sp_el0" while
current_stack_pointer is just "sp" (though I don't think we ever use
current_pt_regs in interrupt context).

-- 
Catalin

Re: [PATCH] arm64: add architecture specified current_pt_regs

[Zhi-zhou]

On Thu, Feb 18, 2016 at 7:58 PM, Catalin Marinas
<catalin.marinas@arm.com> wrote:
>
> On Thu, Feb 18, 2016 at 07:48:35PM +0800, Zhi-zhou Zhang wrote:
> > From: zhizhou <zhizhou.zh@gmail.com>
> >
> > This patch is based on the implementation of arm. The generic
> > current_pt_regs is implemented with current->stack. It need to access
> > memory that would be too expensive.
>
> Do you have any performance numbers?

I'm using QEMU, so no. Actually this macro isn't heavily used. I just
think using the generic
implementation is not very nice. It get task_struct from sp_el0, then
get stack(which is
equal to sp_el0) from task_struct. There are two unnecessary memory accesses.
>
>
> > diff --git a/arch/arm64/include/asm/ptrace.h b/arch/arm64/include/asm/ptrace.h
> > index e9e5467..1865d54 100644
> > --- a/arch/arm64/include/asm/ptrace.h
> > +++ b/arch/arm64/include/asm/ptrace.h
> > @@ -185,5 +185,9 @@ static inline int valid_user_regs(struct user_pt_regs *regs)
> >
> >  extern unsigned long profile_pc(struct pt_regs *regs);
> >
> > +#define current_pt_regs(void) ({ (struct pt_regs *)                  \
> > +     ((current_stack_pointer | (THREAD_SIZE - 1)) - 0xf) - 1;        \
> > +})
>
> I don't think this works well with the separate IRQ stack that we merged
> in 4.5-rc1. current_thread_info() explicitly uses "sp_el0" while
> current_stack_pointer is just "sp" (though I don't think we ever use
> current_pt_regs in interrupt context).
>
Thank you, that's a problem. So how about write it like this?

diff --git a/arch/arm64/include/asm/ptrace.h b/arch/arm64/include/asm/ptrace.h
index e9e5467..b68e01c 100644
--- a/arch/arm64/include/asm/ptrace.h
+++ b/arch/arm64/include/asm/ptrace.h
@@ -185,5 +185,10 @@ static inline int valid_user_regs(struct
user_pt_regs *regs)

 extern unsigned long profile_pc(struct pt_regs *regs);

+#define current_pt_regs(void) ({ (struct pt_regs *)            \
+               (((unsigned long)current_thread_info() |        \
+               (THREAD_SIZE - 1)) - 0xf) - 1;                  \
+})
+
 #endif /* __ASSEMBLY__ */
 #endif

> --
> Catalin




-- 
Regards,
Zhi-zhou

Re: [PATCH] arm64: add architecture specified current_pt_regs

[Will Deacon]

On Fri, Feb 19, 2016 at 10:30:09AM +0800, Zhi-zhou wrote:
> On Thu, Feb 18, 2016 at 7:58 PM, Catalin Marinas
> <catalin.marinas@arm.com> wrote:
> >
> > On Thu, Feb 18, 2016 at 07:48:35PM +0800, Zhi-zhou Zhang wrote:
> > > From: zhizhou <zhizhou.zh@gmail.com>
> > >
> > > This patch is based on the implementation of arm. The generic
> > > current_pt_regs is implemented with current->stack. It need to access
> > > memory that would be too expensive.
> >
> > Do you have any performance numbers?
> 
> I'm using QEMU, so no. Actually this macro isn't heavily used. I just
> think using the generic
> implementation is not very nice. It get task_struct from sp_el0, then
> get stack(which is
> equal to sp_el0) from task_struct. There are two unnecessary memory accesses.

I'd much rather use the generic implementation unless there's a compelling
reason not to. "I think it's not very nice" doesn't really cut it for me!

Will

Re: [PATCH] arm64: add architecture specified current_pt_regs

[Zhizhou Zhang]

On Fri, Feb 19, 2016 at 6:32 PM, Will Deacon <will.deacon@arm.com> wrote:
> On Fri, Feb 19, 2016 at 10:30:09AM +0800, Zhi-zhou wrote:
>> On Thu, Feb 18, 2016 at 7:58 PM, Catalin Marinas
>> <catalin.marinas@arm.com> wrote:
>> >
>> > On Thu, Feb 18, 2016 at 07:48:35PM +0800, Zhi-zhou Zhang wrote:
>> > > From: zhizhou <zhizhou.zh@gmail.com>
>> > >
>> > > This patch is based on the implementation of arm. The generic
>> > > current_pt_regs is implemented with current->stack. It need to access
>> > > memory that would be too expensive.
>> >
>> > Do you have any performance numbers?
>>
>> I'm using QEMU, so no. Actually this macro isn't heavily used. I just
>> think using the generic
>> implementation is not very nice. It get task_struct from sp_el0, then
>> get stack(which is
>> equal to sp_el0) from task_struct. There are two unnecessary memory accesses.
>
> I'd much rather use the generic implementation unless there's a compelling
> reason not to. "I think it's not very nice" doesn't really cut it for me!
Refer to memory twice may incur cache eviction. I'm really a newbie to
kernel. Anyway, I have a look at kernel and I'm very curious what's
the 'compelling reason' of 16 architectures implement their own
current_pt_regs?
>
> Will



-- 
Regards,
Zhizhou