Re: [PATCH 2/3] rbtree: update augmented data on the way down in rb_add_augmented_cached()

[Yiwei Lin]

On 2026/09/28 22:05, Peter Zijlstra wrote:
> > And that is *very* close to being generalizable, obviating the need for
> > RBCOMPUTE. Does your LLM see a way to make that happen?
>
> Perhaps by doing something like so?
[...]
>  #define RB_DECLARE_CALLBACKS_MULTI(RBSTATIC, RBNAME,			\
> -			     RBSTRUCT, RBFIELD, RBCOPY, RBCOMPUTE)	\
> +			     RBSTRUCT, RBFIELD, RBCOMPUTE, RBAUG...)	\

Yes, and since RBCOMPUTE is always "reset the fields to the node's own contribution,
then fold in each child". I think we can also generalize it with RBMERGE, by
adding one extra RBINIT:

        static inline void min_vruntime_init(struct sched_entity *se)
        {
                se->min_vruntime = se->vruntime;
                se->min_slice = se->slice;
                se->max_slice = se->slice;
        }

So the original RBCOMPUTE can be generated by the template instead, which allow us to
remove min_vruntime_update.

        static inline bool RBNAME ## _compute(RBSTRUCT *node)                         \
        {                                                                             \
                typeof(node->RBAUGMENTED) aug;                                        \
                RBSTRUCT *child;                                                      \
                                                                                      \
                RBINIT(node, &aug);                                                   \
                if (node->RBFIELD.rb_left) {                                          \
                        child = rb_entry(node->RBFIELD.rb_left, RBSTRUCT, RBFIELD);   \
                        RBMERGE(&aug, &child->RBAUGMENTED);                           \
                }                                                                     \
                if (node->RBFIELD.rb_right) {                                         \
                        child = rb_entry(node->RBFIELD.rb_right, RBSTRUCT, RBFIELD);  \
                        RBMERGE(&aug, &child->RBAUGMENTED);                           \
                }                                                                     \
                if (!memcmp(&aug, &node->RBAUGMENTED, sizeof(aug)))                   \
                        return true;                                                  \
                node->RBAUGMENTED = aug;                                              \
                return false;                                                         \
        }                                                                             \

Base on this I have another idea: can we let augmented data becomes one member and
init/merge just work on its type by value. Taking sched for example:

        struct_group_tagged(sched_aug, aug,
                u64 min_vruntime;
                u64 min_slice;
                u64 max_slice;
        );

So fair.c's se->min_vruntime etc. can stay as they are, we don't have to rely on the
FOR_EACH machinery. Do you think this can be better or do you prefer the field-list form?

Thanks,
Yiwei Lin