[Intel-gfx] [RFC 11/13] cgroup/drm: Introduce weight based drm cgroup control

[Tvrtko Ursulin]

On 22/11/2022 21:29, Tejun Heo wrote:
> On Wed, Nov 09, 2022 at 04:11:39PM +0000, Tvrtko Ursulin wrote:
>> +DRM scheduling soft limits
>> +~~~~~~~~~~~~~~~~~~~~~~~~~~
>> +
>> +Because of the heterogenous hardware and driver DRM capabilities, soft limits
>> +are implemented as a loose co-operative (bi-directional) interface between the
>> +controller and DRM core.
>> +
>> +The controller configures the GPU time allowed per group and periodically scans
>> +the belonging tasks to detect the over budget condition, at which point it
>> +invokes a callback notifying the DRM core of the condition.
>> +
>> +DRM core provides an API to query per process GPU utilization and 2nd API to
>> +receive notification from the cgroup controller when the group enters or exits
>> +the over budget condition.
>> +
>> +Individual DRM drivers which implement the interface are expected to act on this
>> +in the best-effort manner only. There are no guarantees that the soft limits
>> +will be respected.
> 
> Soft limits is a bit of misnomer and can be confused with best-effort limits
> such as memory.high. Prolly best to not use the term.

Are you suggesting "best effort limits" or "best effort <something>"? It 
would sounds good to me if we found the right <something>. Best effort 
budget perhaps?

>> +static bool
>> +__start_scanning(struct drm_cgroup_state *root, unsigned int period_us)
>> +{
>> +	struct cgroup_subsys_state *node;
>> +	bool ok = false;
>> +
>> +	rcu_read_lock();
>> +
>> +	css_for_each_descendant_post(node, &root->css) {
>> +		struct drm_cgroup_state *drmcs = css_to_drmcs(node);
>> +
>> +		if (!css_tryget_online(node))
>> +			goto out;
>> +
>> +		drmcs->active_us = 0;
>> +		drmcs->sum_children_weights = 0;
>> +
>> +		if (node == &root->css)
>> +			drmcs->per_s_budget_ns =
>> +				DIV_ROUND_UP_ULL(NSEC_PER_SEC * period_us,
>> +						 USEC_PER_SEC);
>> +		else
>> +			drmcs->per_s_budget_ns = 0;
>> +
>> +		css_put(node);
>> +	}
>> +
>> +	css_for_each_descendant_post(node, &root->css) {
>> +		struct drm_cgroup_state *drmcs = css_to_drmcs(node);
>> +		struct drm_cgroup_state *parent;
>> +		u64 active;
>> +
>> +		if (!css_tryget_online(node))
>> +			goto out;
>> +		if (!node->parent) {
>> +			css_put(node);
>> +			continue;
>> +		}
>> +		if (!css_tryget_online(node->parent)) {
>> +			css_put(node);
>> +			goto out;
>> +		}
>> +		parent = css_to_drmcs(node->parent);
>> +
>> +		active = drmcs_get_active_time_us(drmcs);
>> +		if (active > drmcs->prev_active_us)
>> +			drmcs->active_us += active - drmcs->prev_active_us;
>> +		drmcs->prev_active_us = active;
>> +
>> +		parent->active_us += drmcs->active_us;
>> +		parent->sum_children_weights += drmcs->weight;
>> +
>> +		css_put(node);
>> +		css_put(&parent->css);
>> +	}
>> +
>> +	ok = true;
>> +
>> +out:
>> +	rcu_read_unlock();
>> +
>> +	return ok;
>> +}
> 
> A more conventional and scalable way to go about this would be using an
> rbtree keyed by virtual time. Both CFS and blk-iocost are examples of this,
> but I think for drm, it can be a lot simpler.

It's well impressive you were able to figure out what I am doing there. 
:) And probably you can see that this is the first time I am attempting 
an algorithm like this one. I think I made it /dtrt/ with a few post/pre 
walks so the right pieces of data propagate correctly.

Are you suggesting a parallel/shadow tree to be kept in the drm 
controller (which would shadow the cgroup hierarchy)? Or something else? 
The mention of rbtree is not telling me much, but I will look into the 
referenced examples. (Although I will refrain from major rework until 
more people start "biting" into all this.)

Also, when you mention scalability you are concerned about multiple tree 
walks I have per iteration? I wasn't so much worried about that, 
definitely not for the RFC, but even in general due relatively low 
frequency of scanning and a good amount of less trivial cost being 
outside the actual tree walks (drm client walks, GPU utilisation 
calculations, maybe more). But perhaps I don't have the right idea on 
how big cgroups hierarchies can be compared to number of drm clients etc.

Regards,

Tvrtko