On Mon, Oct 07, 2019 at 11:26:35AM -0700, Linus Torvalds wrote:
> The good news is that right now x86 is the only architecture that does
> that user_access_begin(), so we don't need to worry about anything
> else. Apparently the ARM people haven't had enough performance
> problems with the PAN bit for them to care.
Take a look at this:
static inline unsigned long raw_copy_from_user(void *to,
const void __user *from, unsigned long n)
{
unsigned long ret;
if (__builtin_constant_p(n) && (n <= 8)) {
ret = 1;
switch (n) {
case 1:
barrier_nospec();
__get_user_size(*(u8 *)to, from, 1, ret);
break;
case 2:
barrier_nospec();
__get_user_size(*(u16 *)to, from, 2, ret);
break;
case 4:
barrier_nospec();
__get_user_size(*(u32 *)to, from, 4, ret);
break;
case 8:
barrier_nospec();
__get_user_size(*(u64 *)to, from, 8, ret);
break;
}
if (ret == 0)
return 0;
}
barrier_nospec();
allow_read_from_user(from, n);
ret = __copy_tofrom_user((__force void __user *)to, from, n);
prevent_read_from_user(from, n);
return ret;
}
That's powerpc. And while the constant-sized bits are probably pretty
useless there as well, note the allow_read_from_user()/prevent_read_from_user()
part. Looks suspiciously similar to user_access_begin()/user_access_end()...
The difference is, they have separate "for read" and "for write" primitives
and they want the range in their user_access_end() analogue. Separating
the read and write isn't a problem for callers (we want them close to
the actual memory accesses). Passing the range to user_access_end() just
might be tolerable, unless it makes you throw up...
