cvs commit: src/sys/sys sx.h

[John Baldwin]

On Wednesday 30 August 2006 16:51, Ruslan Ermilov wrote:
> On Wed, Aug 30, 2006 at 01:20:32PM -0400, John Baldwin wrote:
> > On Wednesday 30 August 2006 05:37, Bruce Evans wrote:
> > > On Tue, 29 Aug 2006, John Baldwin wrote:
> > > 
> > > > jhb         2006-08-29 20:36:33 UTC
> > > >
> > > >  FreeBSD src repository
> > > >
> > > >  Modified files:
> > > >    sys/sys              sx.h
> > > >  Log:
> > > >  The _sx_assert() prototype should exist if either of INVARIANTS or
> > > >  INVARIANT_SUPPORT is defined so you can build a kernel with
> > > >  INVARIANT_SUPPORT, but build a module with just INVARIANTS on.
> > > 
> > > No it shouldn't.  INVARIANT_SUPPORT is a documented prerequisite for
> > > INVARIANTS.  So is the resulting requirements for using INVARIANTS to
> > > create non-modular "modules":  From /sys/conf/NOTES:
> > > 
> > > # The INVARIANT_SUPPORT option makes us compile in support for
> > > # verifying some of the internal structures.  It is a prerequisite for
> > >                                                ^^^^^^^^^^^^^^^^^^^^^^^^
> > > # 'INVARIANTS', as enabling 'INVARIANTS' will make these functions be
> > >    ^^^^^^^^^^^^
> > > # called.  The intent is that you can set 'INVARIANTS' for single
> > > # source files (by changing the source file or specifying it on the
> > > # command line) if you have 'INVARIANT_SUPPORT' enabled.  Also, if you
> > >                                                            ^^^^^^^^^^^^
> > > # wish to build a kernel module with 'INVARIANTS', then adding
> > >    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> > > # 'INVARIANT_SUPPORT' to your kernel will provide all the necessary
> > >    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> > > # infrastructure without the added overhead.
> > >    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> > > 
> > > However, INVARIANTS is a fairly bogus option.  Last time I looked
> > > (long ago) it only controlled a small amount of kernel bloat, and there
> > > are probably many other functions that are defined unconditionally else
> > > modules with INVARIANTS would be more broken.
> > 
> > Err, the text in NOTES is probably wrong then.  The idea is that 
INVARIANTS 
> > enables various assertions.  Similar to how NDEBUG turns off assert() (but 
> > inverted).  The purpose of INVARIANT_SUPPORT is to provide any needed 
> > assertion-supporting code (such as _foo_assert()) macros in the kernel.  
> > Thus, if I want to run any code that uses INVARIANTs, I need to have a 
kernel 
> > built with INVARIANT_SUPPORT.  However, I might build only selected 
portions 
> > of the kernel with INVARIANTS.  For example, I might build none of the 
kernel 
> > with INVARIANTS, but only a kernel module for a device driver being 
developed 
> > or a test kernel module (like my crash.c and crash2.c.).
> > 
> I read what you've written precisely as what was quoted from NOTES.
> I.e., to be able to compile some bit with INVARIANTS, the kernel
> should have been compiled with INVARIANT_SUPPORT.  _sx_asser() is
> part of that support, not a "user" of invatianted code.  If some
> code that uses sx locks needs to be "invarianted", it can be
> compiled with only "options INVARIANTS".
> 
> I fail to see how your commit follows this, however.  What you have
> committed may eventually turn into "if option INVARINANTS is enabled
> when compiling a kernel, automatically enable the INVARIANT_SUPPORT
> option" -- when all INVARIANT_SUPPORT ifdefs will be converted to
> INVARIANT_SUPPORT || INVARIANTS.

No.  The _sx_assert() _function_ is only #ifdef INVARIANT_SUPPORT in 
kern_sx.c.  However, the _sx_assert() function _prototype_ is used both in 
kern_sx.c when compiling, but also by consumers of the sx_assert() macro 
defined in sys/sx.h.  That is, the prototype is needed for both cases, 1 is 
so kern_sx.c has a prototype for the function it declares, this requires 
INVARIANT_SUPPORT; 2 is for all the users of sx_assert() in the tree, which 
map sx_assert() to a call to _sx_assert() if INVARIANTS is defined.

-- 
John Baldwin