[PATCH v3 0/4] rbtree: declare augmented callbacks per field, fix rb_add_augmented_cached() descent

From: Yiwei Lin

Date: Mon Oct 05 2026 - 05:33:01 EST


rbtree_test open-codes the insertion of every flavour of rbtree it
exercises, while the generic rb_add*() helpers have been the way most
users insert nodes for years now. Patches 1 and 4 make the test use the
helpers where the helper does exactly what the test did, so the helpers
themselves get covered.

Patch 2 is Peter's RB_AUG() rework from the v1 thread: each augmented
field is described by its per-node value, its aggregate member and how
two aggregates combine, and the template derives every callback from
that, per field and in a local; RB_DECLARE_CALLBACKS_MULTI() goes
away, RB_DECLARE_CALLBACKS_MAX() becomes a one-line wrapper and its
RBCOMPUTE is renamed RBVALUE. One change on top of what was posted:
the third argument is a fold(a, b) rather than a "replace?" compare,
which keeps the same register-local recompute, lets sums and counts be
expressed, and replaces RB_MIN/RB_MAX with the existing min()/max().

Converting the cached augmented test exposed the cost of the
"suboptimal" propagate-from-parent path in rb_add_augmented_cached(),
2-12% on the augmented insert+delete benchmark against the documented
update-on-the-way-down pattern. Patch 3 adds a ->merge() callback to
struct rb_augment_callbacks, derived from the RB_AUG() list, and uses
it during the descent, which gets the helper to parity before the test
starts relying on it.

checkpatch has plenty to say about the RB_FOR_EACH() machinery in
patches 2 and 3 (unused macro arguments, values without parentheses);
all of it is inherent to the token-pasting dispatch, as for __MAP() in
linux/syscalls.h. The tools/ copy of rbtree_augmented.h is left alone.

v2: https://lore.kernel.org/r/20260929152439.91443-1-s921975628@xxxxxxxxx
v1: https://lore.kernel.org/r/20260928122611.336351-1-s921975628@xxxxxxxxx

Changes since v2:
- Keep the 'exit' early return of the old RB_DECLARE_CALLBACKS_MAX()
recompute in the per-field recompute

Changes since v1:
- New patch 2: Peter's per-field RB_AUG() templates, taking a fold
- ->merge() is derived from the RB_AUG() list instead of being a new
template argument; the propagate-from-parent removal is now patch 3
- Numbers re-measured on the final code, on the Pi and on x86

Peter Zijlstra (Intel) (1):
rbtree: declare augmented callbacks per field with RB_AUG()

Yiwei Lin (3):
rbtree_test: use rb_add() and rb_add_cached() for the basic tests
rbtree: update augmented data on the way down in
rb_add_augmented_cached()
rbtree_test: use rb_add_augmented_cached() for the cached augmented
test

Documentation/core-api/rbtree.rst | 27 ++++-
include/linux/rbtree_augmented.h | 194 +++++++++++++++++++++---------
kernel/sched/fair.c | 70 ++---------
lib/rbtree_test.c | 62 ++--------
net/tipc/name_table.c | 7 +-
5 files changed, 181 insertions(+), 179 deletions(-)

--
2.34.1