Re: [PATCH] time: Skip overflow itimer tests on 32-bit systems

[Stafford Horne <shorne@gmail.com>]

On Thu, Jun 10, 2021 at 06:38:04AM +0900, Stafford Horne wrote:
> On Wed, Jun 09, 2021 at 10:50:23AM -0300, Adhemerval Zanella wrote:
> > 
> > 
> > On 07/06/2021 10:18, Stafford Horne wrote:
> > > On the port of OpenRISC I am working on and it appears the rv32 port
> > > we have sets __TIMESIZE == 64 && __WORDSIZE == 32.  This causes the
> > > size of time_t to be 8 bytes, but the tv_sec in the kernel is still 32-bit
> > > causing truncation.
> > > 
> > > The truncations are unavoidable on these systems so skip the
> > > testing/failures by guarding with __KERNEL_OLD_TIMEVAL_MATCHES_TIMEVAL64.
> > 
> > Sigh, I was hoping that we won't need to handle this situation (glibc 
> > support only 64-bit time_t, but kernel still providing some 32-bit 
> > syscall).
> > 
> > > ---
> > > 
> > > I am open to other suggestions, this seemed the most correct to me.
> > > 
> > > Cc: Adhemerval Zanella <adhemerval.zanella@linaro.org>
> > > Cc: Alistair Francis <alistair.francis@wdc.com>
> > > 
> > >  time/tst-itimer.c | 2 ++
> > >  1 file changed, 2 insertions(+)
> > > 
> > > diff --git a/time/tst-itimer.c b/time/tst-itimer.c
> > > index 929c2b74c7..0c99d46d7e 100644
> > > --- a/time/tst-itimer.c
> > > +++ b/time/tst-itimer.c
> > > @@ -89,6 +89,7 @@ do_test (void)
> > >        TEST_COMPARE (it.it_interval.tv_sec, it_old.it_interval.tv_sec);
> > >        TEST_COMPARE (it.it_interval.tv_usec, it_old.it_interval.tv_usec);
> > >  
> > > +#if __KERNEL_OLD_TIMEVAL_MATCHES_TIMEVAL64
> > >        if (sizeof (time_t) == 4)
> > >  	continue;
> > >  
> > > @@ -146,6 +147,7 @@ do_test (void)
> > >  	  TEST_COMPARE (setitimer (timers[i], &it, NULL), -1);
> > >  	  TEST_COMPARE (errno, EOVERFLOW);
> > >  	}
> > > +#endif
> > >    }
> > >  
> > >    {
> > > 
> > 
> > Instead of disabling, I think it would be better to use
> > __KERNEL_OLD_TIMEVAL_MATCHES_TIMEVAL64 instead of __time_t sizeof
> > (so we can still tests the EOVERFLOW):
> > 
> > diff --git a/time/tst-itimer.c b/time/tst-itimer.c
> > index 929c2b74c7..bd7d7afe83 100644
> > --- a/time/tst-itimer.c
> > +++ b/time/tst-itimer.c
> > @@ -100,7 +100,7 @@ do_test (void)
> > 
> >        /* Linux does not provide 64 bit time_t support for getitimer and
> >          setitimer on architectures with 32 bit time_t support.  */
> > -      if (sizeof (__time_t) == 8)
> > +      if (__KERNEL_OLD_TIMEVAL_MATCHES_TIMEVAL64)
> >         { 
> >           TEST_COMPARE (setitimer (timers[i], &it, NULL), 0);
> >           TEST_COMPARE (setitimer (timers[i], &(struct itimerval) { 0 },
> > @@ -131,7 +131,7 @@ do_test (void)
> >        it.it_interval.tv_usec = 20;
> >        it.it_value.tv_sec = 30;
> >        it.it_value.tv_usec = 40;
> > -      if (sizeof (__time_t) == 8)
> > +      if (__KERNEL_OLD_TIMEVAL_MATCHES_TIMEVAL64)
> >         { 
> >           TEST_COMPARE (setitimer (timers[i], &it, NULL), 0);
> 
> This looks good to me, I can update to this, test and resend the patch when I
> get some time.  Probably later tonight.

I tested this and it exposes an issue in the linux setitimer wrapper.  On my
platform I get EINVAL instead of EOVERFLOW.

    FAIL: time/tst-itimer
    original exit status 1
    tst-itimer.c:125: numeric comparison failure
       left: 22 (0x16); from: errno
      right: 75 (0x4b); from: EOVERFLOW
    tst-itimer.c:147: numeric comparison failure
       left: 22 (0x16); from: errno
      right: 75 (0x4b); from: EOVERFLOW


It seems this is because sysdeps/unix/sysv/linux/setitimer.c, checks that
the incoming value is in the range of time_t.  The problem is that
that we need to fit the value in __int32_t not time_t.  When testing the time_t
range check does not detect the overflow and setitimer ends up passing a -1 to
the kernel causing EINVAL.

I can fix that, as per the patch below, but It will take me some time to audit
other places this might be an issue.



  if (! in_time_t_range (new_value->it_interval.tv_sec)
      || ! in_time_t_range (new_value->it_value.tv_sec))
    {
      __set_errno (EOVERFLOW);
      return -1;
    }
  new_value_32.it_interval
    = valid_timeval64_to_timeval32 (new_value->it_interval);
  new_value_32.it_value
    = valid_timeval64_to_timeval32 (new_value->it_value);




The below patch works for me, but there is probably a better thing to do then
create a new functrion.



diff --git a/include/time.h b/include/time.h
index 4372bfbd96..377a4a45ea 100644
--- a/include/time.h
+++ b/include/time.h
@@ -342,6 +342,14 @@ in_time_t_range (__time64_t t)
   return s == t;
 }
 
+/* Check whether T fits in a timeval32 (__int32_t).  */
+static inline bool
+in_timeval32_range (__time64_t t)
+{
+  __int32_t s = t;
+  return s == t;
+}
+
 /* Convert a known valid struct timeval into a struct __timespec64.  */
 static inline struct __timespec64
 valid_timeval_to_timespec64 (const struct timeval tv)
diff --git a/sysdeps/unix/sysv/linux/setitimer.c
b/sysdeps/unix/sysv/linux/setitimer.c
index 083a25cf35..bada30ba02 100644
--- a/sysdeps/unix/sysv/linux/setitimer.c
+++ b/sysdeps/unix/sysv/linux/setitimer.c
@@ -32,8 +32,8 @@ __setitimer64 (__itimer_which_t which,
 #else
   struct __itimerval32 new_value_32;
 
-  if (! in_time_t_range (new_value->it_interval.tv_sec)
-      || ! in_time_t_range (new_value->it_value.tv_sec))
+  if (! in_timeval32_range (new_value->it_interval.tv_sec)
+      || ! in_timeval32_range (new_value->it_value.tv_sec))
     {


-Stafford