[PATCH v6 3/3] aarch64: tests for SME

[Yury Khrustalev]

Hi,

Thanks for the suggestions for improving the tests!

On Tue, Oct 07, 2025 at 01:28:36PM -0300, Adhemerval Zanella Netto wrote:
> 
> On 06/10/25 08:53, Yury Khrustalev wrote:
> > This commit adds tests for the following use cases relevant to handing of
> > the SME state:
> > 
> >  - fork() and vfork()
> >  - clone() and clone3()
> >  - signal handler
> > 
> > ...
> >
> > +  /* Enabled ZA state so that effect of disabling be observable.  */
> > +  enable_sme_za_state (ptr);
> > +  check_sme_za_state ("before clone", /* Clear.  */ false);
> > +
> > +  pid_t pid = __clone3 (&args, sizeof (args), fun, syscall_name);
> 
> Unfortunately this does not work on all scenarios.  With gcc 15.2.1 with
> glibc configured with --enable-stack-protector=all I see:
> 
> $ make test t=misc/tst-sme-clone3
> [...]
> syscall error 22 (Invalid argument)
> make[2]: Leaving directory '/home/azanella/Projects/glibc/glibc-git/misc'
> FAIL: misc/tst-sme-clone3

Good catch.

> ... 
> I think it would be better to pull the implementation from glibc itself,
> which is the one that we should actually test:

Right. I went this way originally, but after undefined symbol errors I
thought there would be more, so opted out for using the libgcc
implementation.

> -$(objpfx)tst-sme-clone3: $(objpfx)clone3.o
> +$(objpfx)tst-sme-clone3: $(objpfx)clone3.o \
> +                        $(objpfx)__arm_za_disable.o

Right

> +/* Required by __arm_za_disable.S and provided by statup code as hidden
> +   symbol.  */
> +uint64_t _dl_hwcap2;

A neat trick, thanks.

> ... 
> > +
> > +  /* Check that ZA state of SME was disabled in parent.  */
> > +  check_sme_za_state ("after clone in parent", /* Clear.  */ true);
> > +
> > +  printf ("%s child pid: %d\n", syscall_name, pid);
> > +  if (pid == -1)
> > +    {
> > +      if (errno == ENOSYS)
> > +        {
> > +          puts ("clone3 syscall is not supported");
> > +          exit (EXIT_UNSUPPORTED);
> > +        }
> > +      perror ("clone3");
> > +      TEST_VERIFY_EXIT (0);
> > +    }
> 
> I think we can assume that HWCAP_SME (Linux 5.19) implies in clone3 (Linux 5.3)
> support. 

I'm dropping this part: as you correctly pointed out below, we'd exit on
any syscall error anyway.

> > +  if (waitid (P_PID, pid, NULL, WEXITED))
> 
> Maybe xwaitpid?

Agreed.

> > +/* Workaround to simplify linking with clone3.o.  */
> > +void __syscall_error(int code)
> > +{
> > +  int err = -code;
> > +  fprintf (stderr, "syscall error %d (%s)\n", err, strerror (err));
> > +  exit (err);
> 
> This makes all the clone3 error handling superflous.

Correct.

> > ...
> >
> > +static void
> > +handler (int signo, siginfo_t *si, void *ctx)
> > +{
> > +  TEST_VERIFY (signo == SIGUSR1);
> > +  printf ("in the handler: %d and %p\n", signo, si);
> 
> The printf here is not async-signal-safe, but I think it should not matter
> for this test.

We can use xwrite() here if we don't print signo and si (which is rather
unnecessary at this point).

> > ...
> >
> > +  enable_sme_za_state (blk);
> > +  check_sme_za_state ("before signal", false /* State is not clear.  */);
> > +  xraise (SIGUSR1);
> 
> Afaik SIGUSR1 is not a synchronous signal, meaning that the next 
> check_sme_za_state state is not guarantee to execute *after* the signal handler
> is being acted uppon.
> 
> The best strategy would do something like:
> 
> ...
> 
> +static int pipefd[2];
> +

Another neat trick, thanks.

> > ...
> >
> > +  pid_t pid = vfork ();
> > +
> > +  if (pid == 0)
> > +    {
> > +      /* Check that ZA state of SME was disabled in child.  */
> > +      check_sme_za_state ("after vfork in child", /* Clear.  */ true);
> > +      exit (0);
> 
> Use _exit.  It is undefined calling any other function other than
> execve/_exit after vfork; but I think it should not matter on this
> test.

Good point.

> ...

Thanks, I'll be sending v7 shortly.

Kind regards,
Yury