Adrian mentioned a few weeks ago that KCONFIG_ALLCONFIG is the way to
go to ensure that things like allyes/allmod/allnoconfig work with a
constrained set of symbols, with the implication that this holds
true for randconfig as well.
While allyes/mod/noconfigs do seem to work fine with KCONFIG_ALLCONFIG
provisions, randconfig tramples all over the provided values at perhaps
not surprisingly, random.
Debugging this a bit, there seemed to be two issues:
- SYMBOL_DEF and SYMBOL_DEF_USER overlap, which made
def_sym->flags the same regardless of whether we came from an
KCONFIG_ALLCONFIG path or not.
- clobbering of the fixed value in conf_choice() by way of
random() def assignment.
While I don't pretend to have any idea what I'm doing in Kconfig,
changing SYMBOL_DEF_USER to reuse the SYMBOL_DEF3 bit (and thereby
separating it from SYMBOL_DEF) in addition to checking the def_sym flag
before randomly assigning a new value ended up fixing things up.
Signed-off-by: Paul Mundt <[email protected]>
---
scripts/kconfig/conf.c | 9 ++++++++-
scripts/kconfig/expr.h | 3 +--
2 files changed, 9 insertions(+), 3 deletions(-)
diff --git a/scripts/kconfig/conf.c b/scripts/kconfig/conf.c
index a38787a..61a61df 100644
--- a/scripts/kconfig/conf.c
+++ b/scripts/kconfig/conf.c
@@ -374,7 +374,14 @@ static int conf_choice(struct menu *menu)
continue;
break;
case set_random:
- def = (random() % cnt) + 1;
+ /*
+ * Defined symbols need to be honoured even in
+ * the randomization case, or else symbols set
+ * via the KCONFIG_ALLCONFIG path are clobbered.
+ * - PFM.
+ */
+ if (!(def_sym->flags & SYMBOL_DEF))
+ def = (random() % cnt) + 1;
case set_default:
case set_yes:
case set_mod:
diff --git a/scripts/kconfig/expr.h b/scripts/kconfig/expr.h
index a195986..f5b4873 100644
--- a/scripts/kconfig/expr.h
+++ b/scripts/kconfig/expr.h
@@ -95,9 +95,8 @@ struct symbol {
#define SYMBOL_CHECKED 0x2000
#define SYMBOL_WARNED 0x8000
#define SYMBOL_DEF 0x10000
-#define SYMBOL_DEF_USER 0x10000
#define SYMBOL_DEF_AUTO 0x20000
-#define SYMBOL_DEF3 0x40000
+#define SYMBOL_DEF_USER 0x40000
#define SYMBOL_DEF4 0x80000
#define SYMBOL_MAXLENGTH 256
Hi,
On Wed, 28 Nov 2007, Paul Mundt wrote:
> Adrian mentioned a few weeks ago that KCONFIG_ALLCONFIG is the way to
> go to ensure that things like allyes/allmod/allnoconfig work with a
> constrained set of symbols, with the implication that this holds
> true for randconfig as well.
BTW another possibility is to use all{no,mod,yes,random,}.config.
> While allyes/mod/noconfigs do seem to work fine with KCONFIG_ALLCONFIG
> provisions, randconfig tramples all over the provided values at perhaps
> not surprisingly, random.
Please be careful with such broad statements, there is only an issue with
choice values.
> Debugging this a bit, there seemed to be two issues:
>
> - SYMBOL_DEF and SYMBOL_DEF_USER overlap, which made
> def_sym->flags the same regardless of whether we came from an
> KCONFIG_ALLCONFIG path or not.
Look at how SYMBOL_DEF is used in confdata.c.
> - clobbering of the fixed value in conf_choice() by way of
> random() def assignment.
Simply add a test for is_new there.
bye, Roman
On Wed, Nov 28, 2007 at 06:08:16PM +0100, Roman Zippel wrote:
> On Wed, 28 Nov 2007, Paul Mundt wrote:
> > While allyes/mod/noconfigs do seem to work fine with KCONFIG_ALLCONFIG
> > provisions, randconfig tramples all over the provided values at perhaps
> > not surprisingly, random.
>
> Please be careful with such broad statements, there is only an issue with
> choice values.
>
Ok, I'll rephrase, '100% of the provided values I tested with were being
randomly clobbered'. Is that better? Broken is broken, whether it applies
to a small subset of symbols or not.
> > Debugging this a bit, there seemed to be two issues:
> >
> > - SYMBOL_DEF and SYMBOL_DEF_USER overlap, which made
> > def_sym->flags the same regardless of whether we came from an
> > KCONFIG_ALLCONFIG path or not.
>
> Look at how SYMBOL_DEF is used in confdata.c.
>
Ah, ok. I was just trying to find something I could test that would be
different for the KCONFIG_ALLCONFIG path, but it seems like is_new is a
much cleaner solution for this, thanks for pointing it out!
Updated patch follows.
Signed-off-by: Paul Mundt <[email protected]>
---
scripts/kconfig/conf.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/scripts/kconfig/conf.c b/scripts/kconfig/conf.c
index a38787a..8d6f174 100644
--- a/scripts/kconfig/conf.c
+++ b/scripts/kconfig/conf.c
@@ -374,7 +374,8 @@ static int conf_choice(struct menu *menu)
continue;
break;
case set_random:
- def = (random() % cnt) + 1;
+ if (is_new)
+ def = (random() % cnt) + 1;
case set_default:
case set_yes:
case set_mod: