On Fri, Mar 19, 2021 at 10:24:03PM +0300, Dmitry Osipenko wrote:

19.03.2021 22:03, Minchan Kim пишет:

quoted

On Fri, Mar 19, 2021 at 09:48:11PM +0300, Dmitry Osipenko wrote:

quoted

19.03.2021 21:21, Minchan Kim пишет:

quoted

On Fri, Mar 19, 2021 at 08:56:06PM +0300, Dmitry Osipenko wrote:

quoted

19.03.2021 19:30, Minchan Kim пишет:

quoted

+static void cma_kobj_release(struct kobject *kobj)
+{
+	struct cma_kobject *cma_kobj = container_of(kobj, struct cma_kobject, kobj);
+
+	kfree(cma_kobj);
+}

Oh, wait.. I think this kfree wrong since cma_kobj belongs to the array.

Oh, good spot. Let me use kzalloc.

Thinking a bit more about this.. it looks like actually it should be
better to get back to the older variant of cma_stat, but allocate at the
time of CMA initialization, rather than at the time of sysfs
initialization. Then the cma_stat will be decoupled from the cma struct

IIRC, the problem was slab was not initiaized at CMA init point.
That's why I liked your suggestion.

Alright, if CMA init time is a problem, then the recent variant should
be okay.

quoted

quoted

and cma_stat will be a self-contained object.

Yeah, self-contained is better but it's already weird to
have differnt lifetime for one object since CMA object
never die, technically.

Indeed.

I found the Greg's original argument and not sure that it's really
worthwhile to worry about the copycats since this is not a driver's code..

Maybe we could just add a clarifying comment for the kobj, telling why
it's okay for CMA. Greg, doesn't it sound like a good compromise to you?

Please no.