Re: [PATCH 2/4] Add NR_syscalls for x86_64

[Jason Baron <jbaron@redhat.com>]

On Tue, Aug 25, 2009 at 03:04:15PM -0400, Mathieu Desnoyers wrote:
> * Jason Baron (jbaron@redhat.com) wrote:
> > On Tue, Aug 25, 2009 at 09:40:05AM -0400, Jason Baron wrote:
> > > > > On Mon, Aug 24, 2009 at 05:40:18PM -0400, Jason Baron wrote:
> > > > > > Add a dynamic definition of NR_syscalls for x86_64. This is used in the 
> > > > > > syscall events tracing code.
> > > > > > 
> > > > > > Signed-off-by: Jason Baron <jbaron@redhat.com>
> > > > > > 
> > > > > > ---
> > > > > >  include/linux/ftrace.h |    5 +++++
> > > > > >  1 files changed, 5 insertions(+), 0 deletions(-)
> > > > > > 
> > > > > > diff --git a/include/linux/ftrace.h b/include/linux/ftrace.h
> > > > > > index dc3b132..3bd94ee 100644
> > > > > > --- a/include/linux/ftrace.h
> > > > > > +++ b/include/linux/ftrace.h
> > > > > > @@ -14,6 +14,11 @@
> > > > > >  
> > > > > >  #include <asm/ftrace.h>
> > > > > >  
> > > > > > +#ifdef CONFIG_X86_64
> > > > > > +#include <asm/asm-offsets.h>
> > > > > > +#define NR_syscalls (__NR_syscall_max + 1)
> > > > > > +#endif
> > > > > > +
> > > > > 
> > > > > 
> > > > > I'd rather see this in arch/x86/include/unistd_64.h, since
> > > > > NR_syscalls seems to be the defined in unistd.h in all archs.
> > > > > 
> > > > > That also makes a unified way to retrieve this number for
> > > > > other users.
> > > > > 
> > > > 
> > > > Yes, NR_syscalls should be included there.
> > > > 
> > > > -- Steve
> > > 
> > > right, for x86_64, unistd.h is included to generate __NR_syscall_max
> > > which is then used for NR_syscalls. So I did initially try it there, but
> > > there were dependency problems. I'll see what I can come up with...
> > > 
> > > thanks,
> > > 
> > > -Jason
> > 
> > ok, since unistd_64.h is used to generate offsets.c, which eventually
> > generates, asm/asm-offsets.h, which defines, __NR_syscall_max, we can't
> > just put #define of NR_syscalls in unistd_64.h. However, it seems the
> > only point in time that unistd_64.h can not depend upon __NR_syscall_max,
> > is before asm/asm-offsets.h is defined. Thus, the patch below passes in
> > a '-DCREATE_OFFSETS' during the creation of asm/asm-offsets.h, so that
> > we do not include it before it exists. The patch is hacky but works.
> > thoughts?
> 
> Ugh! My eyes hurt!
> 

sorry :)

> What you are doing here is to basically put back the hardcoded
> NR_syscalls rather that using the build infrastructure already in place.
> 

no. NR_syscalls is not hardcoded by this patch. Its defined in terms of
__NR_syscall_max which is dynamically generated by the kernel build.

> If my memory serves me well, unistd_64.h generates __NR_syscall_max
> automatically by being included multiples times. Can we generalize this
> and make the information generated available in an automaticaly
> generated header instead ? It is saved in ams-offsets.h currently as
> "__NR_syscall_max". We could also save it somewhere else meant to be
> included by C code.
> 
> Mathieu
> 

The request was to define NR_syscalls in unistd.h, since that is the
historical Linux location for it. Adding another automatically generated
header does not accomplish that. Even if I include that new file in
unistd.h, I'm still going to have a circular dependency, and require a
solution similar to what I've proposed.

thanks,

-Jason