Currently there are four places that end up calling into __xa_cmpxchg().
Two in drivers/infiniband/hw/mlx5/odp.c, one in fs/erofs/utils.c and one
in mm/shmem.c. The xarray in the first three places do not contain any
marks ever, the fourth place originally used radix_tree_delete_item()
which didn't touch marks either. So we are safe to switch __xa_cmpxchg()
to use xas_store_noinit(). Also document that __xa_cmpxchg() now does
not touch marks at all - i.e., it has ordinary store semantics without
specialcasing NULL value. If someone ever needs xa_cmpxchg() equivalent
that would really cause mark clearing on storing NULL, we can create
xa_erase_item() function.
Signed-off-by: Jan Kara <jack@xxxxxxx>
---
lib/xarray.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/lib/xarray.c b/lib/xarray.c
index a372c59e3914..d87045d120ad 100644
--- a/lib/xarray.c
+++ b/lib/xarray.c
@@ -1460,7 +1460,8 @@ EXPORT_SYMBOL(xa_store);
*
* You must already be holding the xa_lock when calling this function.
* It will drop the lock if needed to allocate memory, and then reacquire
- * it afterwards.
+ * it afterwards. The call does not change any xarray marks except for
+ * XA_FREE_MARK if free tracking is enabled.
*
* Context: Any context. Expects xa_lock to be held on entry. May
* release and reacquire xa_lock if @gfp flags permit.
@@ -1478,7 +1479,7 @@ void *__xa_cmpxchg(struct xarray *xa, unsigned long index,
do {
curr = xas_load(&xas);
if (curr == old) {
- xas_store(&xas, entry);
+ xas_store_noinit(&xas, entry);
if (xa_track_free(xa)) {
if (entry && !curr)
xas_clear_mark(&xas, XA_FREE_MARK);
--
2.16.4
