On 2024-12-03 08:58:26-0500, James Bottomley wrote:

On Tue, 2024-12-03 at 21:02 +0800, Zijun Hu wrote:

quoted

On 2024/12/3 20:41, Greg Kroah-Hartman wrote:

quoted

On Tue, Dec 03, 2024 at 08:23:45PM +0800, Zijun Hu wrote:

[...]

quoted

quoted

quoted

or squash such patch series into a single patch ?

various subsystem maintainers may not like squashing way.

Agreed, so look into either doing it in a bisectable way if at all
possible.  As I don't see a full series here, I can't suggest how
it needs to happen :(

let me send you a full series later and discuss how to solve this
issue.

It's only slightly more complex than what we normally do: modify all
instances and then change the API.  In this case you have an additional
problem because the prototype "const void *" will cause a mismatch if a
function has "void *".  The easiest way to solve this is probably to
make device_find_child a macro that coerces its function argument to
having a non const "void *" and then passes off to the real function. 
If you do that in the first patch, then you can constify all the
consumers and finally remove the macro coercion in the last patch.

Casting function pointers like that should be detected and trapped by
control flow integrity checking (KCFI).

Another possibility would be to use a macro and _Generic to dispatch to
two different backing functions. See __BIN_ATTR() in
include/linux/sysfs.h for an inspiration.
This also enables an incremental migration.


Thomas