On Fri, 15 Nov 2024 at 15:06, Josh Poimboeuf [off-list ref] wrote:
So I think the thing to do is

 (a) find out which __get_user() it is that matters so much for that load

Do you have a profile somewhere?

 (b) convert them to use "unsafe_get_user()", with that whole

                if (can_do_masked_user_access())
                        from = masked_user_access_begin(from);
                else if (!user_read_access_begin(from, sizeof(*from)))
                        return -EFAULT;

     sequence before it.

And if it's just a single __get_user() (rather than a sequence of
them), just convert it to get_user().

Hmm?

The profile is showing futex_get_value_locked():

int futex_get_value_locked(u32 *dest, u32 __user *from)
{
	int ret;

	pagefault_disable();
	ret = __get_user(*dest, from);
	pagefault_enable();

	return ret ? -EFAULT : 0;
}

That has several callers, so we can probably just use get_user() there?

Also, is there any harm in speeding up __get_user()?  It still has ~80
callers and it's likely to be slowing down things we don't know about.

It's usually only the regressions which get noticed, and that LFENCE
went in almost 7 years ago, when there was much less automated
performance regression testing.

As a bonus, that patch will root out any "bad" users, which will
eventually allow us to simplify things and just make __get_user() an
alias of get_user().

In fact, if we aliased it for all arches, that could help in getting rid
of __get_user() altogether as there would no longer be any (real or
advertised) benefit to using it.

-- 
Josh