On Wed, Dec 11, 2024 at 04:32:50AM +0000, Matthew Wilcox (Oracle) wrote:
[...]
> +int obj_cgroup_charge_vmalloc(struct obj_cgroup **objcgp,
> + unsigned int nr_pages, gfp_t gfp)
> +{
> + struct obj_cgroup *objcg;
> + int err;
> +
> + if (mem_cgroup_disabled() || !(gfp & __GFP_ACCOUNT))
> + return 0;
> +
> + objcg = current_obj_cgroup();
> + if (!objcg)
> + return 0;
> +
> + err = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
> + if (err)
> + return err;
> + obj_cgroup_get(objcg);
> + mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_VMALLOC, nr_pages);
obj_cgroup_memcg() needs to be within rcu. See MEMCG_PERCPU_B as an
example.
> + *objcgp = objcg;
> +
> + return 0;
> +}
> +
> +/**
> + * obj_cgroup_uncharge_vmalloc - Uncharge vmalloc memory
> + * @objcg: The object cgroup
> + * @nr_pages: Number of pages
> + */
> +void obj_cgroup_uncharge_vmalloc(struct obj_cgroup *objcg,
> + unsigned int nr_pages)
> +{
> + if (!objcg)
> + return;
> + mod_memcg_state(objcg->memcg, MEMCG_VMALLOC, 0L - nr_pages);
Please use obj_cgroup_memcg() above instead of objcg->memcg and within
rcu lock.
Overall the patch looks good.
