Re: [PATCH v4 2/2] of: overlay: rework overlay apply and remove kfree()s

From: Frank Rowand
Date: Wed Apr 20 2022 - 18:31:45 EST


On 4/20/22 17:25, frowand.list@xxxxxxxxx wrote:
> From: Frank Rowand <frank.rowand@xxxxxxxx>
>

< snip >

The changes between v3 and v4, via interdiff:


diff -u b/drivers/of/overlay.c b/drivers/of/overlay.c
--- b/drivers/of/overlay.c
+++ b/drivers/of/overlay.c
@@ -60,10 +60,10 @@
* @new_fdt: Memory allocated to hold unflattened aligned FDT
* @overlay_mem: the memory chunk that contains @overlay_root
* @overlay_root: expanded device tree that contains the fragment nodes
+ * @notify_state: most recent notify action used on overlay
* @count: count of fragment structures
* @fragments: fragment nodes in the overlay expanded device tree
* @symbols_fragment: last element of @fragments[] is the __symbols__ node
- * @kfree_unsafe: pointers into the @new_fdt or @overlay_mem may exist
* @cset: changeset to apply fragments to live device tree
*/
struct overlay_changeset {
@@ -72,10 +72,10 @@
const void *new_fdt;
const void *overlay_mem;
struct device_node *overlay_root;
+ enum of_overlay_notify_action notify_state;
int count;
struct fragment *fragments;
bool symbols_fragment;
- bool kfree_unsafe;
struct of_changeset cset;
};

@@ -165,6 +165,8 @@
struct of_overlay_notify_data nd;
int i, ret;

+ ovcs->notify_state = action;
+
for (i = 0; i < ovcs->count; i++) {
struct fragment *fragment = &ovcs->fragments[i];

@@ -864,17 +866,17 @@
* There should be no live pointers into ovcs->overlay_mem and
* ovcs->new_fdt due to the policy that overlay notifiers are not
* allowed to retain pointers into the overlay devicetree other
- * than during the window between OF_OVERLAY_PRE_APPLY overlay
- * notifiers and the OF_OVERLAY_POST_REMOVE overlay notifiers.
- * During the window, ovcs->kfree_unsafe will be true.
+ * than during the window from OF_OVERLAY_PRE_APPLY overlay
+ * notifiers until the OF_OVERLAY_POST_REMOVE overlay notifiers.
*
- * A memory leak will occur here if ovcs->kfree_unsafe is true.
+ * A memory leak will occur here if within the window.
*/

- if (!ovcs->kfree_unsafe)
+ if (ovcs->notify_state == OF_OVERLAY_INIT ||
+ ovcs->notify_state == OF_OVERLAY_POST_REMOVE) {
kfree(ovcs->overlay_mem);
- if (!ovcs->kfree_unsafe)
kfree(ovcs->new_fdt);
+ }
kfree(ovcs);
}

@@ -902,10 +904,8 @@
* following attempt to apply or remove an overlay changeset will be
* refused.
*
- * Returns 0 on success, or a negative error number. When references to
- * ovcs->new_fdt and ovcs->overlay_root may exist, ovcs->kfree_unsafe is
- * set to true. On error return, the caller of of_overlay_apply() must
- * call free_overlay_changeset().
+ * Returns 0 on success, or a negative error number. On error return,
+ * the caller of of_overlay_apply() must call free_overlay_changeset().
*/

static int of_overlay_apply(struct overlay_changeset *ovcs)
@@ -926,12 +926,6 @@
if (ret)
goto out;

- /*
- * After overlay_notify(), ovcs->overlay_root related pointers may have
- * leaked to drivers, so can not kfree() ovcs->overlay_mem and
- * ovcs->new_fdt until after OF_OVERLAY_POST_REMOVE notifiers.
- */
- ovcs->kfree_unsafe = true;
ret = overlay_notify(ovcs, OF_OVERLAY_PRE_APPLY);
if (ret) {
pr_err("overlay changeset pre-apply notify error %d\n", ret);
@@ -1001,6 +995,11 @@
of_overlay_mutex_lock();
mutex_lock(&of_mutex);

+ /*
+ * ovcs->notify_state must be set to OF_OVERLAY_INIT before allocating
+ * ovcs resources, implicitly set by kzalloc() of ovcs
+ */
+
ovcs->id = idr_alloc(&ovcs_idr, ovcs, 1, 0, GFP_KERNEL);
if (ovcs->id <= 0) {
ret = ovcs->id;
only in patch2:
unchanged:
--- a/include/linux/of.h
+++ b/include/linux/of.h
@@ -1543,7 +1543,8 @@ static inline bool of_device_is_system_power_controller(const struct device_node
*/

enum of_overlay_notify_action {
- OF_OVERLAY_PRE_APPLY = 0,
+ OF_OVERLAY_INIT = 0, /* kzalloc() of ovcs sets this value */
+ OF_OVERLAY_PRE_APPLY,
OF_OVERLAY_POST_APPLY,
OF_OVERLAY_PRE_REMOVE,
OF_OVERLAY_POST_REMOVE,