x86/uaccess: Avoid barrier_nospec() in 64-bit copy_from_user()

The barrier_nospec() in 64-bit copy_from_user() is slow. Instead use
pointer masking to force the user pointer to all 1's for an invalid
address.

The kernel test robot reports a 2.6% improvement in the per_thread_ops
benchmark [1].

This is a variation on a patch originally by Josh Poimboeuf [2].

Link: https://lore.kernel.org/[e-mail] [1]
Link: https://lore.kernel.org/[e-mail] [2]
Tested-and-reviewed-by: Josh Poimboeuf <[e-mail]>
Cc: Kirill A. Shutemov <[e-mail]>
Signed-off-by: Linus Torvalds <[e-mail]>

include/linux/uaccess.h

···
 #else
  #define can_do_masked_user_access() 0
  #define masked_user_access_begin(src) NULL
+ #define mask_user_address(src) (src)
 #endif
 
 /*
···
 {
 	unsigned long res = n;
 	might_fault();
-	if (!should_fail_usercopy() && likely(access_ok(from, n))) {
+	if (should_fail_usercopy())
+		goto fail;
+	if (can_do_masked_user_access())
+		from = mask_user_address(from);
+	else {
+		if (!access_ok(from, n))
+			goto fail;
 		/*
 		 * Ensure that bad access_ok() speculation will not
 		 * lead to nasty side effects *after* the copy is
 		 * finished:
 		 */
 		barrier_nospec();
-		instrument_copy_from_user_before(to, from, n);
-		res = raw_copy_from_user(to, from, n);
-		instrument_copy_from_user_after(to, from, n, res);
 	}
-	if (unlikely(res))
-		memset(to + (n - res), 0, res);
+	instrument_copy_from_user_before(to, from, n);
+	res = raw_copy_from_user(to, from, n);
+	instrument_copy_from_user_after(to, from, n, res);
+	if (likely(!res))
+		return 0;
+fail:
+	memset(to + (n - res), 0, res);
 	return res;
 }
 extern __must_check unsigned long