Re: [PATCH v5 02/36] compiler-context-analysis: Add infrastructure for Context Analysis with Clang

[Marco Elver <elver@google.com>]

On Fri, Dec 19, 2025 at 11:04AM -0800, 'Bart Van Assche' via kasan-dev wrote:
> On 12/19/25 10:59 AM, Marco Elver wrote:
> > On Fri, 19 Dec 2025 at 19:39, 'Bart Van Assche' via kasan-dev
> > <kasan-dev@googlegroups.com> wrote:
> > > I'm concerned that the context_lock_struct() macro will make code harder
> > > to read. Anyone who encounters the context_lock_struct() macro will have
> > > to look up its definition to learn what it does. I propose to split this
> > > macro into two macros:
> > > * One macro that expands into "__ctx_lock_type(name)".
> > > * A second macro that expands into the rest of the above macro.
> > > 
> > > In other words, instead of having to write
> > > context_lock_struct(struct_name, { ... }); developers will have to write
> > > 
> > > struct context_lock_type struct_name {
> > >       ...;
> > > };
> > > context_struct_helper_functions(struct_name);
> > 
> > This doesn't necessarily help with not having to look up its
> > definition to learn what it does.
> > 
> > If this is the common pattern, it will blindly be repeated, and this
> > adds 1 more line and makes this a bit more verbose. Maybe the helper
> > functions aren't always needed, but I also think that context lock
> > types should remain relatively few.  For all synchronization
> > primitives that were enabled in this series, the helpers are required.
> > 
> > The current usage is simply:
> > 
> > context_lock_struct(name) {
> >     ... struct goes here ...
> > };  // note no awkward ) brace
> > 
> > I don't know which way the current kernel style is leaning towards,
> > but if we take <linux/cleanup.h> as an example, a simple programming
> > model / API is actually preferred.
> Many kernel developers are used to look up the definition of a data
> structure either by using ctags, etags or a similar tool or by using
> grep and a pattern like "${struct_name} {\$". Breaking the tools kernel
> developer use today to look up data structure definitions might cause
> considerable frustration and hence shouldn't be done lightly.

Fair point. In fact, it's as simple as e.g. (just tested with mutex) as
this:

diff --git a/include/linux/mutex_types.h b/include/linux/mutex_types.h
index 80975935ec48..63ab9e65bb48 100644
--- a/include/linux/mutex_types.h
+++ b/include/linux/mutex_types.h
@@ -38,7 +38,8 @@
  * - detects multi-task circular deadlocks and prints out all affected
  *   locks and tasks (and only those tasks)
  */
-context_lock_struct(mutex) {
+context_lock_struct(mutex);
+struct mutex {
 	atomic_long_t		owner;
 	raw_spinlock_t		wait_lock;
 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
@@ -59,7 +60,8 @@ context_lock_struct(mutex) {
  */
 #include <linux/rtmutex.h>
 
-context_lock_struct(mutex) {
+context_lock_struct(mutex);
+struct mutex {
 	struct rt_mutex_base	rtmutex;
 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 	struct lockdep_map	dep_map;

So the existing macro does support both use-cases as-is. I suppose we
could force the above use pattern.

The reason it works, is because it forward-declares the struct anyway to
define the helper functions.