[PATCH 2/2] arm64: Avoid memcpy() for syscall_get_arguments()

[david laight]

On Mon, 1 Dec 2025 10:13:54 +0000
Mark Rutland <mark.rutland at arm.com> wrote:

> On Thu, Nov 27, 2025 at 08:36:30PM +0800, Jinjie Ruan wrote:
> > Do not use memcpy() to extract syscall arguments from struct pt_regs
> > but rather just perform direct assignments.
> > 
> > The performance benchmarks with Generic Entry patch[1] with audit on
> > from perf bench basic syscall on kunpeng920 gives roughly a 1%
> > performance uplift and also aligns the implementation with
> > x86 and RISC-V.
> > 
> > | Metric     | W/O this patch | With this patch | Change    |
> > | ---------- | -------------- | --------------- | --------- |
> > | Total time | 2.241 [sec]    | 2.211 [sec]     |  ↓1.36%   |
> > | usecs/op   | 0.224157       | 0.221146        |  ↓1.36%   |
> > | ops/sec    | 4,461,157      | 4,501,409       |  ↑0.9%    |
> > 
> > Before:
> > <syscall_get_arguments.constprop.0>:
> >        aa0103e2        mov     x2, x1
> >        91002003        add     x3, x0, #0x8
> >        f9408804        ldr     x4, [x0, #272]
> >        f8008444        str     x4, [x2], #8
> >        a9409404        ldp     x4, x5, [x0, #8]
> >        a9009424        stp     x4, x5, [x1, #8]
> >        a9418400        ldp     x0, x1, [x0, #24]
> >        a9010440        stp     x0, x1, [x2, #16]
> >        f9401060        ldr     x0, [x3, #32]
> >        f9001040        str     x0, [x2, #32]
> >        d65f03c0        ret
> >        d503201f        nop
> > 
> > After:
> >        a9408e82        ldp     x2, x3, [x20, #8]
> >        2a1603e0        mov     w0, w22
> >        f9400e84        ldr     x4, [x20, #24]
> >        f9408a81        ldr     x1, [x20, #272]
> >        9401c4ba        bl      ffff800080215ca8 <__audit_syscall_entry>  
> 
> It's probably worth noting that __audit_syscall_entry() only takes 4
> syscall arguments, and hence the compiler has elided the copy of
> regs->regs[4] and regs->regs[5], which it apparently couldn't manage
> before.

Hasn't it actually inlined it and completely optimised away the regs[] array?
It looks (from the asm) as though syscall_get_arguments() is followed by:
	fn(regs[0], regs[1], regs[2], regs[3])

    David

> 
> > [1]: https://lore.kernel.org/all/20251126071446.3234218-1-ruanjinjie@huawei.com/
> > Signed-off-by: Jinjie Ruan <ruanjinjie at huawei.com>
> > ---
> >  arch/arm64/include/asm/syscall.h | 8 +++++---
> >  1 file changed, 5 insertions(+), 3 deletions(-)
> > 
> > diff --git a/arch/arm64/include/asm/syscall.h b/arch/arm64/include/asm/syscall.h
> > index f3853047c28e..f3564ba97f7e 100644
> > --- a/arch/arm64/include/asm/syscall.h
> > +++ b/arch/arm64/include/asm/syscall.h
> > @@ -82,9 +82,11 @@ static inline void syscall_get_arguments(struct task_struct *task,
> >  					 unsigned long *args)
> >  {
> >  	args[0] = regs->orig_x0;
> > -	args++;
> > -
> > -	memcpy(args, &regs->regs[1], 5 * sizeof(args[0]));
> > +	args[1] = regs->regs[1];
> > +	args[2] = regs->regs[2];
> > +	args[3] = regs->regs[3];
> > +	args[4] = regs->regs[4];
> > +	args[5] = regs->regs[5];
> >  }  
> 
> FWIW, I think this is clearer than the 'args++' and the memcpy(), so I'm
> happy with this regardless of the performance concern.
> 
> However, as Dmitry says, we should keep this structurally the same as
> syscall_set_arguments(), and so we should update that in the same way.
> 
> Mark.
>