Re: V2 [PATCH 2/2] Add a syscall test for [BZ #25810]

["H.J. Lu" <hjl.tools@gmail.com>]

On Wed, Apr 22, 2020 at 5:47 AM Florian Weimer <fw@deneb.enyo.de> wrote:
>
> * H. J. Lu:
>
> > On Wed, Apr 22, 2020 at 5:25 AM Florian Weimer <fw@deneb.enyo.de> wrote:
> >>
> >> * H. J. Lu via Libc-alpha:
> >>
> >> > Add a test to pass 64-bit long arguments to syscall with undefined upper
> >> > 32 bits on ILP32 systems.
> >>
> >> What does this test, exactly?  How does it ensure that the upper bits
> >> are indeed non-zero, to exercise the zero-extension case?
> >
> > On x32,
> >
> > struct Array
> > {
> >   size_t length;
> >   void *ptr;
> > };
> >
> > can be passed in a single 64-bit integer register.  When a 32-bit
> > integer is passed in
> > a 64-bit integer, its upper 32 bits can contain undefined value:
> >
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94541
> >
> > This testcase arranges syscalls in such a way that the upper 32 bits
> > of 64 big integer
> > register, which is used to pass unsigned long to kernel, contains the
> > "ptr" value passed in
> > function arguments.   If the upper 32 bits aren't cleared, syscall
> > will fail since long in kernel
> > is 64 bits, not 32 bits.
>
> Would you please add this as a comment to the patch, and one-line
> comments where the clobbers happen?

I will add

+/* On x32, this can be passed in a single 64-bit integer register.  */
 struct Array
 {
   size_t length;
@@ -50,6 +51,9 @@ __attribute__ ((noclone, noinline))
 void
 deallocate (struct Array b)
 {
+  /* On x32, the 64-bit integer register containing `b' may be copied
+     to another 64-bit integer register to pass the second argument to
+     munmap.  */
   if (b.length && munmap (b.ptr, b.length))
     {
       printf ("munmap error: %m\n");

> And say that the test is in this formulation likely x32-specific, but
> that it is expected to work on other architectures as well.



-- 
H.J.