Re: [PATCH 1/1] PATCH: KGDB/KDB Fix no KDB config problem.

[Ingo Molnar <mingo@kernel.org>]

* Mike Travis <travis@sgi.com> wrote:

> 
> 
> On 10/3/2013 10:15 AM, Ingo Molnar wrote:
> > 
> > * Mike Travis <travis@sgi.com> wrote:
> > 
> >>
> >>
> >> On 10/3/2013 9:51 AM, Ingo Molnar wrote:
> >>>
> >>> * Mike Travis <travis@sgi.com> wrote:
> >>>
> >>>> Some code added to the debug_core module had KDB dependencies
> >>>> that it shouldn't have.
> >>>>
> >>>> Signed-off-by: Mike Travis <travis@sgi.com>
> >>>> ---
> >>>>  kernel/debug/debug_core.c |    8 ++++----
> >>>>  kernel/debug/debug_core.h |    2 ++
> >>>>  2 files changed, 6 insertions(+), 4 deletions(-)
> >>>>
> >>>> --- linux.orig/kernel/debug/debug_core.c
> >>>> +++ linux/kernel/debug/debug_core.c
> >>>> @@ -575,8 +575,8 @@ return_normal:
> >>>>  		raw_spin_lock(&dbg_slave_lock);
> >>>>  
> >>>>  #ifdef CONFIG_SMP
> >>>> -	/* If SYSTEM_NMI, slaves are already waiting */
> >>>> -	if (ks->err_code == KDB_REASON_SYSTEM_NMI)
> >>>> +	/* If send_ready set, slaves are already waiting */
> >>>> +	if (ks->send_ready)
> >>>>  		atomic_set(ks->send_ready, 1);
> >>>>  
> >>>>  	/* Signal the other CPUs to enter kgdb_wait() */
> >>>> @@ -682,11 +682,11 @@ kgdb_handle_exception(int evector, int s
> >>>>  	if (arch_kgdb_ops.enable_nmi)
> >>>>  		arch_kgdb_ops.enable_nmi(0);
> >>>>  
> >>>> +	memset(ks, 0, sizeof(struct kgdb_state));
> >>>>  	ks->cpu			= raw_smp_processor_id();
> >>>>  	ks->ex_vector		= evector;
> >>>>  	ks->signo		= signo;
> >>>>  	ks->err_code		= ecode;
> >>>> -	ks->kgdb_usethreadid	= 0;
> >>>>  	ks->linux_regs		= regs;
> >>>>  
> >>>>  	if (kgdb_reenter_check(ks))
> >>>> @@ -750,7 +750,7 @@ int kgdb_nmicallin(int cpu, int trapnr,
> >>>>  		ks->cpu			= cpu;
> >>>>  		ks->ex_vector		= trapnr;
> >>>>  		ks->signo		= SIGTRAP;
> >>>> -		ks->err_code		= KDB_REASON_SYSTEM_NMI;
> >>>> +		ks->err_code		= KGDB_KDB_REASON_SYSTEM_NMI;
> >>>>  		ks->linux_regs		= regs;
> >>>>  		ks->send_ready		= send_ready;
> >>>>  		kgdb_cpu_enter(ks, regs, DCPU_WANT_MASTER);
> >>>> --- linux.orig/kernel/debug/debug_core.h
> >>>> +++ linux/kernel/debug/debug_core.h
> >>>> @@ -75,11 +75,13 @@ extern int kdb_stub(struct kgdb_state *k
> >>>>  extern int kdb_parse(const char *cmdstr);
> >>>>  extern int kdb_common_init_state(struct kgdb_state *ks);
> >>>>  extern int kdb_common_deinit_state(void);
> >>>> +#define KGDB_KDB_REASON_SYSTEM_NMI KDB_REASON_SYSTEM_NMI
> >>>>  #else /* ! CONFIG_KGDB_KDB */
> >>>>  static inline int kdb_stub(struct kgdb_state *ks)
> >>>>  {
> >>>>  	return DBG_PASS_EVENT;
> >>>>  }
> >>>> +#define KGDB_KDB_REASON_SYSTEM_NMI 0
> >>>>  #endif /* CONFIG_KGDB_KDB */
> >>>>  
> >>>>  #endif /* _DEBUG_CORE_H_ */
> >>>
> >>> Hm, the KGDB_KDB_REASON_SYSTEM_NMI definition is a bit ugly. I still think 
> >>> there are layering violations here and just kludging it around doesn't 
> >>> solve it - a helper function that keeps kgdb details to the kgdb code 
> >>> would.
> >>>
> >>> Anyway, if Jason is fine with this solution and upholds his Acked-by then 
> >>> I'll merge this into the first patch and apply the two patches.
> >>>
> >>> Thanks,
> >>>
> >>> 	Ingo
> >>>
> >>
> >> Would you prefer a simple #ifdef CONFIG_KGDB_KDB around the assignment?
> > 
> > An #ifdef doesn't solve the layering violation!
> > 
> > kernel/debug/debug_core.c didn't have any serious use of #ifdef 
> > CONFIG_KGDB_KDB before (it used it to flavor a few printks) and I'm not 
> > convinced it needs one for this feature either.
> > 
> > So if that whole chunk is kdb dependent, why not shuffle that into a 
> > nicely named function and host it somewhere appropriate in 
> > kernel/debug/kdb/? That function would turn into an empty inline function 
> > in the !CONFIG_KGDB_KDB case.
> > 
> > But ... it's up to Jason really whether he wants to abstract that piece of 
> > code out, I'm just kibitzing here really.
> > 
> > Thanks,
> > 
> > 	Ingo
> > 
> 
> If I moved it to KDB then we would lose the capability to enter KGDB 
> mode and the gdb_stub functionality.  The #ifdef could have a comment to 
> this effect?

No, this really suggests that there's something else broken around there.

Removing #ifdefs around non-trivial code and architecting it properly 
really makes things more reviewable, easier to understand, better 
specified and generally less bug prone. So adding an #ifdef around a pair 
of printk()s is probably OK - anything that is longer than a line probably 
deserves abstracted out.

Thanks,

	Ingo