From: Al Viro
> Sent: 26 March 2021 16:12
>
> On Fri, Mar 26, 2021 at 03:38:30PM +0100, Christoph Hellwig wrote:
>
> > +static const char __user *
> > +get_user_arg_ptr(const char __user *const __user *argv, int nr)
> > {
> > + if (in_compat_syscall()) {
> > + const compat_uptr_t __user *compat_argv =
> > + compat_ptr((unsigned long)argv);
> > compat_uptr_t compat;
> >
> > + if (get_user(compat, compat_argv + nr))
> > return ERR_PTR(-EFAULT);
> > return compat_ptr(compat);
> > + } else {
> > + const char __user *native;
> >
> > + if (get_user(native, argv + nr))
> > + return ERR_PTR(-EFAULT);
> > + return native;
> > + }
> > }
>
> Yecchhh.... So you have in_compat_syscall() called again and again, for
> each argument in the list? I agree that current version is fucking ugly,
> but I really hate that approach ;-/
Especially since in_compat_syscall() isn't entirely trivial on x86-64.
Probably all in the noise for 'exec', but all the bits do add up.
You may not want separate get_user() on some architectures either.
The user_access_begin/end aren't cheap.
OTOH if you call copy_from_user() you get hit by the stupid
additional costs of 'user copy hardening'.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
