dtrace discuss - Apr 2006 - Problems including Sun supplied header files

Howdy,

While attempting to access the rnode_t type and the VTOR() macro in 
/usr/include/nfs/rnode.h, I bumped into another preprocessor issue:

$ uname -a
SunOS winnie 5.11 opensol-b37 sun4u sparc SUNW,Ultra-5_10

$ dtrace -C -s ick.d
"/dev/fd/4", line 2: #include of /usr/include/... may be non-portable
dtrace: failed to compile script ick.d: "/usr/include/nfs/rnode.h",
line
164: syntax error near "kthread_t"

$ cat ick.d
#include "/usr/include/nfs/rnode.h"

fbt:nfs:nfs3_access:entry
{
}

Anyone happen to know what is causing this? Since rnode.h is a Sun 
supplied header file, I would think it should just work.

Thanks,
- Ryan
--
UNIX Administrator
http://daemons.net/~matty

On Sun, Apr 16, 2006 at 11:45:50PM -0400, Matty wrote:
> While attempting to access the rnode_t type and the VTOR() macro in 
> /usr/include/nfs/rnode.h, I bumped into another preprocessor issue:
> 
> $ uname -a
> SunOS winnie 5.11 opensol-b37 sun4u sparc SUNW,Ultra-5_10
> 
> $ dtrace -C -s ick.d
> "/dev/fd/4", line 2: #include of /usr/include/... may be
non-portable
> dtrace: failed to compile script ick.d:
"/usr/include/nfs/rnode.h", line
> 164: syntax error near "kthread_t"
Try including <sys/thread.h>...

# cat /tmp/ick.d
#include <sys/thread.h>
#include <nfs/rnode.h>

fbt:nfs:nfs3_access:entry
{
}
# dtrace -C -s /tmp/ick.d
dtrace: script ''/tmp/ick.d'' matched 1 probe
^C

# 
> Anyone happen to know what is causing this? Since rnode.h is a Sun 
> supplied header file, I would think it should just work.
Perhaps <nfs/rnode.h> or <nfs/nfs.h> ought to include
<sys/thread.h>...

Five headers in /usr/include/nfs/ reference kthread_t.

Nico
--

This doesn''t seem to be a DTrace issue, and is easily reproduced if you
try to compile a .c file that only includes this header file.  Looks
like it uses ''kthread_t'' but doesn''t bring in
sys/thread.h, presumably
because all the current consumers include it manually beforehand.  Feel
free to file a bug.  In the meantime you can just include <sys/thread.h>
before <nfs/rnode.h>.

Hope that helps,

- Eric

On Sun, Apr 16, 2006 at 11:45:50PM -0400, Matty wrote:
> 
> Howdy,
> 
> While attempting to access the rnode_t type and the VTOR() macro in 
> /usr/include/nfs/rnode.h, I bumped into another preprocessor issue:
> 
> $ uname -a
> SunOS winnie 5.11 opensol-b37 sun4u sparc SUNW,Ultra-5_10
> 
> $ dtrace -C -s ick.d
> "/dev/fd/4", line 2: #include of /usr/include/... may be
non-portable
> dtrace: failed to compile script ick.d:
"/usr/include/nfs/rnode.h", line
> 164: syntax error near "kthread_t"
> 
> $ cat ick.d
> #include "/usr/include/nfs/rnode.h"
> 
> fbt:nfs:nfs3_access:entry
> {
> }
> 
> Anyone happen to know what is causing this? Since rnode.h is a Sun 
> supplied header file, I would think it should just work.
> 
> Thanks,
> - Ryan
> --
> UNIX Administrator
> http://daemons.net/~matty
> _______________________________________________
> dtrace-discuss mailing list
> dtrace-discuss at opensolaris.org
--
Eric Schrock, Solaris Kernel Development       http://blogs.sun.com/eschrock

>While attempting to access the rnode_t type and the VTOR() macro in 
>/usr/include/nfs/rnode.h, I bumped into another preprocessor issue:
Dtrace doesn''t allow including random C .h files.

The rnode_t is known to dtrace because it''s compiled into the kernel
(CTF is used to store compiled type information in the kernel)
>Anyone happen to know what is causing this? Since rnode.h is a Sun 
>supplied header file, I would think it should just work.
If included in a kernel module, yes.  Not in D.

Casper

On Mon, 17 Apr 2006, Casper.Dik at Sun.COM wrote:
>
>> While attempting to access the rnode_t type and the VTOR() macro in
>> /usr/include/nfs/rnode.h, I bumped into another preprocessor issue:
>
> Dtrace doesn''t allow including random C .h files.
I was under the impression that you could use #include to add header files 
with types and macros? A quick search through the "Bin" directory in
the
DTraceToolkit seems to indicate that this is possible:

$ grep include tcpsnoop
#include <sys/file.h>
#include <inet/common.h>
#include <sys/byteorder.h>
#include <sys/socket.h>
#include <sys/socketvar.h>

Can someone from the DTrace team shed some light on this for me? I am 
getting conflicting information.
> The rnode_t is known to dtrace because it''s compiled into the
kernel
> (CTF is used to store compiled type information in the kernel)
>
>> Anyone happen to know what is causing this? Since rnode.h is a Sun
>> supplied header file, I would think it should just work.
>
> If included in a kernel module, yes.  Not in D.
Type information is available, but macro definitions are not. I want to 
use macros from /usr/include/* in my scripts to simplify the 
implementation.

Thanks,
- Ryan
--
UNIX Administrator
http://daemons.net/~matty

On Mon, Apr 17, 2006 at 10:28:28AM -0400, Matty wrote:
> On Mon, 17 Apr 2006, Casper.Dik at Sun.COM wrote:
> 
> >
> >>While attempting to access the rnode_t type and the VTOR() macro in
> >>/usr/include/nfs/rnode.h, I bumped into another preprocessor issue:
> >
> >Dtrace doesn''t allow including random C .h files.
That''s incorrect -- DTrace _does_ allow including C header files, with
the
caveat that there are some constructs that a C compiler can absorb in a
C header file that the D compiler will choke on.  The most obvious of these
is the use of D keywords -- if you have a type named "string" D
won''t
be able to compile it.  Slightly more subtle is the declaration of complete
functions and/or static globals in C header files.  This is an absolutely
terrible practice (absent optimization, the text for the function and/or the
data for the global will exist in every object that includes the file), and
solving it is very hard for D.  (It''s one thing to be able to deal with
C
structure and type definitions, but quite another to be able to parse
arbitrary C.)  So if you run into this issue, you will need to either
(1) find a way to not include the header file or (2) wrap the definition
with "#ifndef __SUNW_D".  Finally, you might run into preprocessor
directives that are not supported by the (ancient) version of cpp that we
run.  (An example of such a directive is ''##'', used
(ironically) in
/usr/include/sys/dtrace.h.)  If you get nailed by this, try GCC''s cpp,
as
outlined on slide 46 of the "Advanced DTrace" presentation:

  http://blogs.sun.com/roller/page/bmc?entry=dtrace_tips_tricks_and_gotchas
> I was under the impression that you could use #include to add header files 
> with types and macros? 
Yes, that''s correct.  And we actually test that we can include every
file in /usr/include/sys that can be successfully compiled by including
it alone.  (As of this writing, there are four exceptions on 64-bit,
five on 32-bit.)  Of course, we wouldn''t have picked up your rnode.h
problem
because (1) it''s in /usr/include/nfs (we should add that to the list)
and
(2) it can''t be compiled by including it alone (and as Eric mentioned,
this
is just a bug in rnode.h).

	- Bryan

--------------------------------------------------------------------------
Bryan Cantrill, Solaris Kernel Development.       http://blogs.sun.com/bmc

>>> While attempting to access the rnode_t type and the VTOR() macro in
>>> /usr/include/nfs/rnode.h, I bumped into another preprocessor issue:
>>
>> Dtrace doesn''t allow including random C .h files.
>
>I was under the impression that you could use #include to add header files 
>with types and macros? A quick search through the "Bin" directory
in the
>DTraceToolkit seems to indicate that this is possible:
Hm, I was not quite expecting that.

And generally for types this is not required.  And since D isn''t
C I expect that:
	- there''s not testing that it''s still valid D syntax
	- it may fail at any time.

>$ grep include tcpsnoop
>#include <sys/file.h>
>#include <inet/common.h>
>#include <sys/byteorder.h>
>#include <sys/socket.h>
>#include <sys/socketvar.h>
Interesting.
>Type information is available, but macro definitions are not. I want to 
>use macros from /usr/include/* in my scripts to simplify the 
>implementation.
Ah,

Casper

On Mon, 17 Apr 2006, Bryan Cantrill wrote:
>> I was under the impression that you could use #include to add header
files
>> with types and macros?
>
> Yes, that''s correct.  And we actually test that we can include
every
> file in /usr/include/sys that can be successfully compiled by including
> it alone.  (As of this writing, there are four exceptions on 64-bit,
> five on 32-bit.)  Of course, we wouldn''t have picked up your
rnode.h problem
> because (1) it''s in /usr/include/nfs (we should add that to the
list) and
> (2) it can''t be compiled by including it alone (and as Eric
mentioned, this
> is just a bug in rnode.h).
Thanks for the clarification. I filed CR #6413818 to address the bug I 
bumped into.

Thanks,
- Ryan
--
UNIX Administrator
http://daemons.net/~matty