Re: [PATCH 08/17] drm/cgroup: Track DRM clients per cgroup

[Tejun Heo]

Hello,

On Wed, Jul 12, 2023 at 12:45:56PM +0100, Tvrtko Ursulin wrote:
> +void drmcgroup_client_migrate(struct drm_file *file_priv)
> +{
> +	struct drm_cgroup_state *src, *dst;
> +	struct cgroup_subsys_state *old;
> +
> +	mutex_lock(&drmcg_mutex);
> +
> +	old = file_priv->__css;
> +	src = css_to_drmcs(old);
> +	dst = css_to_drmcs(task_get_css(current, drm_cgrp_id));
> +
> +	if (src != dst) {
> +		file_priv->__css = &dst->css; /* Keeps the reference. */
> +		list_move_tail(&file_priv->clink, &dst->clients);
> +	}
> +
> +	mutex_unlock(&drmcg_mutex);
> +
> +	css_put(old);
> +}
> +EXPORT_SYMBOL_GPL(drmcgroup_client_migrate);

So, you're implicitly migrating the fd to the latest ioctl user on the first
access. While this may work for state-less control like usage time. This
likely will cause problem if you later want to add stateful control for
memory consumption. e.g. What happens if the new destination doesn't have
enough budget to accommodate the fd's usage? Let's say we allow over-commit
with follow-up reclaim or oom kill actions, if so, can we guarantee that all
memory ownership for can always be updated? If not, what happens after
failure?

If DRM needs to transfer fd ownership with resources attached to it, it
likely would be a good idea to make that an explicit operation so that the
attempt can be verified and failed if necessary.

Thanks.

-- 
tejun