[PATCH RFC 09/12] mm/vmalloc: collapse busy-tree find-then-unlink into a single mas_erase

[Pranjal Arya <pranjal.arya@oss.qualcomm.com>]

find_unlink_vmap_area() previously walked the busy tree once (via
find_vmap_area_busy_locked() / __find_vmap_area_mt) and then erased
via a second walker (unlink_vmap_area_busy_locked).  Two independent
maple-tree descents, one per call.

maple_tree exposes mas_erase() which combines lookup and erase in
one descent.  Replace the find+unlink pair with a single mas_erase()
walker.

Signed-off-by: Pranjal Arya <pranjal.arya@oss.qualcomm.com>
---
 mm/vmalloc.c | 18 ++++++++++++++++--
 1 file changed, 16 insertions(+), 2 deletions(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 1b73001e197e..463127d5ce58 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -3122,10 +3122,24 @@ static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
 	do {
 		vn = &vmap_nodes[i];
 
+		/*
+		 * Combine the lookup and removal into a single maple-tree
+		 * descent. mas_erase() positions the state at @addr and clears
+		 * the slot in one pass, returning the previously stored VA.
+		 * This saves the second mas_store(NULL) the original
+		 * find_vmap_area_busy_locked + unlink_vmap_area_busy_locked
+		 * pair issued, halving the busy-tree maple work per vfree.
+		 */
 		spin_lock(&vn->busy.lock);
-		va = find_vmap_area_busy_locked(addr, vn);
+		if (likely(vn->busy.mt_enabled)) {
+			MA_STATE(mas, &vn->busy.mt, addr, addr);
+
+			va = mas_erase(&mas);
+		} else {
+			va = NULL;
+		}
 		if (va)
-			unlink_vmap_area_busy_locked(va, vn);
+			INIT_LIST_HEAD(&va->list);
 		spin_unlock(&vn->busy.lock);
 
 		if (va)

-- 
2.34.1