Hi,
I am trying to transform Linux kernel code in order to test for data
races. Let me illustrate this by an example
struct lpctx {
spinlock_t lock;
int cnt;
};
static struct lpctx *lpctx;
static void lpctx_good1(struct lpctx *lx)
{
spin_lock(&lx->lock);
lx->cnt++;
spin_unlock(&lx->lock);
}
static void lpctx_warning1(struct lpctx *lx)
{
lx->cnt++;
}
As you can see lptctx_warning1() is increasing the counter without
holding the lock. The main idea(*) is to annotate all accesses to lx by
adding a WARN_ON(!lock_is_held(&lx->lock.dep_map)).
Thanks to Julia's excellent help on IRC I have currently this somewhat
working cocci script:
@depends on patch@
struct lpctx *x;
identifier f;
statement S;
@@
( S
+ WARN_ON(!lock_is_held(&x->lock.dep_map));
&
x->f
)
This results in:
@@ -16,10 +16,13 @@ static void lpctx_good1(struct lpctx *lx
DBG("");
spin_lock(&lx->lock);
+ WARN_ON(!lock_is_held(&lx->lock.dep_map));
lx->cnt++;
+ WARN_ON(!lock_is_held(&lx->lock.dep_map));
spin_unlock(&lx->lock);
+ WARN_ON(!lock_is_held(&lx->lock.dep_map));
}
static void lpctx_warning1(struct lpctx *lx)
@@ -27,6 +30,7 @@ static void lpctx_warning1(struct lpctx
DBG("");
lx->cnt++;
+ WARN_ON(!lock_is_held(&lx->lock.dep_map));
}
There are a bunch of problems, e,g. the hard coded type/names or the
WARN_ON() should be put in front of the deferences but that is something
for later to improve. The main problem I face at this point is to filter
out the spin_lock() and spin_unlock() access. All my attempts didn't let
to the expected result. I think the best thing would be to match on the
type. So if 'f' is of type spinlock_t ignore it. Any ideas how this
could be expressed?
Thanks,
Daniel
(*) Talking with Nicholas over a coffee helps a lot. Thanks a lot!
_______________________________________________
Cocci mailing list
[email protected]
https://systeme.lip6.fr/mailman/listinfo/cocci
