dtrace discuss - Jul 2006 - args typed to ''struct timeval'' leads to invalid alignment error

Hi,

I have following script:

=================#!/usr/sbin/dtrace -Fs

/*
 struct timeval {
        long tv_sec;
        long tv_usec;
 };

clnt_cots_kcallit(CLIENT *h, rpcproc_t procnum, xdrproc_t xdr_args,
  caddr_t argsp, xdrproc_t xdr_results, caddr_t resultsp, struct timeval 
wait)
*/


::clnt_cots_kcallit:entry
{
 printf("wait=%d",args[6].tv_sec);
}

::clnt_cots_kcallit:entry
{
 printf("wait=%d",arg6);
}

::clnt_cots_kcallit:entry
{
 printf("wait=%llx",arg6);
}
==================================

Output is like this:


dtrace: error on enabled probe ID 1 (ID 55082: 
fbt:rpcmod:clnt_cots_kcallit:entry): invalid alignment (0x3c) in action 
#1 at DIF offset 8    << Probe1
  0  -> clnt_cots_kcallit                     wait=60  << Probe2
  0   | clnt_cots_kcallit:entry               wait=3c  << Probe3


The timeval value is <60 sec; 0 usec>.  Probe 2 shows the sec correctly.
Probe1 gives error invalid alignment (0x3c). Ox3c is 60 decimal - it 
means Probe1 was taking value 60 as an address.

According to docs.sun.com: Solaris Dynamic Tracing Guide the args[] 
should be typed.
For our args[6] it should be typed to struct timeval, but is not.
Question is whether this is problem in implementation or in 
documentation or whether this is correct behaviour.

ANother question - What is the most simple access to the timeval.tv_sec 
and timeval.tv_usec values?

Thank you,

Pavel Filipensky

On Wed, Jul 12, 2006 at 05:49:18PM +0200, pavel filipensky - Sun Microsystems -
Prague Czech Republic wrote:
> dtrace: error on enabled probe ID 1 (ID 55082: 
> fbt:rpcmod:clnt_cots_kcallit:entry): invalid alignment (0x3c) in action 
> #1 at DIF offset 8    << Probe1
>  0  -> clnt_cots_kcallit                     wait=60  << Probe2
>  0   | clnt_cots_kcallit:entry               wait=3c  << Probe3
> 
> 
> The timeval value is <60 sec; 0 usec>.  Probe 2 shows the sec
correctly.
> Probe1 gives error invalid alignment (0x3c). Ox3c is 60 decimal - it 
> means Probe1 was taking value 60 as an address.
> 
> According to docs.sun.com: Solaris Dynamic Tracing Guide the args[] 
> should be typed.
> For our args[6] it should be typed to struct timeval, but is not.
> Question is whether this is problem in implementation or in 
> documentation or whether this is correct behaviour.
It looks like the implementation is incorrect;  it''s assuming that,
since the
type is a structure, the passed argument is a pointer value.  That''s
correct for large structures, but for for small structures, the ABI specifies 
that they are passed as multiple arguments to the function.  In this case,
arg6 is the tv_sec field, and arg7 is the tv_usec field.

Cheers,
- jonathan
> ANother question - What is the most simple access to the timeval.tv_sec 
> and timeval.tv_usec values?
arg6 and arg7, respectively.

Cheers,
- jonathan

-- 
Jonathan Adams, Solaris Kernel Development

On Wed, Jul 12, 2006 at 09:52:02AM -0700, Jonathan Adams
wrote:
> On Wed, Jul 12, 2006 at 05:49:18PM +0200, pavel filipensky - Sun
Microsystems - Prague Czech Republic wrote:
> > dtrace: error on enabled probe ID 1 (ID 55082: 
> > fbt:rpcmod:clnt_cots_kcallit:entry): invalid alignment (0x3c) in
action
> > #1 at DIF offset 8    << Probe1
> >  0  -> clnt_cots_kcallit                     wait=60  <<
Probe2
> >  0   | clnt_cots_kcallit:entry               wait=3c  << Probe3
> > 
> > 
> > The timeval value is <60 sec; 0 usec>.  Probe 2 shows the sec
correctly.
> > Probe1 gives error invalid alignment (0x3c). Ox3c is 60 decimal - it 
> > means Probe1 was taking value 60 as an address.
> > 
> > According to docs.sun.com: Solaris Dynamic Tracing Guide the args[] 
> > should be typed.
> > For our args[6] it should be typed to struct timeval, but is not.
> > Question is whether this is problem in implementation or in 
> > documentation or whether this is correct behaviour.
> 
> It looks like the implementation is incorrect;  it''s assuming
that, since the
> type is a structure, the passed argument is a pointer value. 
That''s
> correct for large structures, but for for small structures, the ABI
specifies
> that they are passed as multiple arguments to the function.  In this case,
> arg6 is the tv_sec field, and arg7 is the tv_usec field.
I''ve filed:

6448142 args[] handling doesn''t work for small structures

to cover this.

Cheers,
- jonathan

-- 
Jonathan Adams, Solaris Kernel Development

Jonathan Adams wrote:
> On Wed, Jul 12, 2006 at 09:52:02AM -0700, Jonathan Adams wrote:
>> It looks like the implementation is incorrect;  it''s assuming
that, since the
>> type is a structure, the passed argument is a pointer value. 
That''s
>> correct for large structures, but for for small structures, the ABI
specifies
>> that they are passed as multiple arguments to the function.  In this
case,
>> arg6 is the tv_sec field, and arg7 is the tv_usec field.
> 
> I''ve filed:
> 
> 6448142 args[] handling doesn''t work for small structures
Erk!  A trifle unfortunate that users of dtrace need to know
details of the ABI to that level.  Is it in theory possible
for Dtrace to hide that?

- Jeremy

On Thu, Jul 13, 2006 at 09:30:15AM +0100, Jeremy Harris
wrote:
> Jonathan Adams wrote:
> >On Wed, Jul 12, 2006 at 09:52:02AM -0700, Jonathan Adams wrote:
> >>It looks like the implementation is incorrect;  it''s
assuming that, since
> >>the
> >>type is a structure, the passed argument is a pointer value. 
That''s
> >>correct for large structures, but for for small structures, the ABI
> >>specifies that they are passed as multiple arguments to the
function.  In
> >>this case,
> >>arg6 is the tv_sec field, and arg7 is the tv_usec field.
> >
> >I''ve filed:
> >
> >6448142 args[] handling doesn''t work for small structures
> 
> Erk!  A trifle unfortunate that users of dtrace need to know
> details of the ABI to that level.  Is it in theory possible
> for Dtrace to hide that?
Yes.  It''s just work that hasn''t been done yet.

Cheers,
- jonathan

-- 
Jonathan Adams, Solaris Kernel Development