Re: [PATCH 2/3] x86: Pass in pt_regs pointer for syscalls that need it

[Tejun Heo <htejun@gmail.com>]

Ingo Molnar wrote:
> * Tejun Heo <htejun@gmail.com> wrote:
> 
>> Brian Gerst wrote:
>>> x86-64 doesn't have the tail-call problem because it doesn't use the
>>> pt_regs on stack trick for syscall args.  All the args are passed in
>>> registers.
>> Yeah, I was saying that we can do about the same thing on x86_32 by
>> passing in pointer to pt_regs and defining proper syscall wrappers.
>> It will cost a bit of performance by increasing register pressure tho.
> 
> Do you mean converting:
> 
> ptregscall int sys_execve(struct pt_regs *regs, char __user *u_filename,
>                           char __user * __user *argv,
>                           char __user * __user *envp)
> 
> to:
> 
> ptregscall int sys_execve(struct pt_regs *regs)
> {
> 	char __user *u_filename		= syscall_arg1(regs);
> 	char __user * __user *argv	= syscall_arg2(regs);
> 	char __user * __user *envp	= syscall_arg3(regs);
> 
> etc.?

Not exactly.  include/linux/syscalls.h already has syscall wrapping
macros defined, with slight modification to allow archs to define its
own __SC_DECL and __SC_LONG (probably should use different name tho),
the outer function can be easily defined to take pt_regs pointer and
pass in the correct argument to the actual implementation function.
The only added overhead would be pt_regs pointer having to be loaded
into %edi and it having to stay somewhere in the callee till the last
parameter access.

Thanks.

-- 
tejun