[Intel-gfx] [RFC 6/8] drm/i915: Expose per-engine client busyness

[Chris Wilson]

Quoting Tvrtko Ursulin (2020-01-10 14:09:09)
> 
> On 10/01/2020 13:58, Chris Wilson wrote:
> > Quoting Tvrtko Ursulin (2020-01-10 13:30:47)
> >> +static ssize_t
> >> +show_client_busy(struct device *kdev, struct device_attribute *attr, char *buf)
> >> +{
> >> +       struct i915_engine_busy_attribute *i915_attr =
> >> +               container_of(attr, typeof(*i915_attr), attr);
> >> +       struct list_head *list = &i915_attr->client->ctx_list;
> >> +       unsigned int engine_class = i915_attr->engine_class;
> >> +       struct i915_gem_context *ctx;
> >> +       u64 total = 0;
> >> +
> >> +       if (i915_attr->no_busy_stats)
> >> +               return -ENODEV;
> >> +
> >> +       rcu_read_lock();
> >> +       list_for_each_entry_rcu(ctx, list, client_link)
> >> +               total += sw_busy_add(ctx, engine_class);
> >> +       rcu_read_unlock();
> >> +
> >> +       return snprintf(buf, PAGE_SIZE, "%llu\n", total);
> >> +}
> >> +
> >> +static const char *uabi_class_names[] = {
> >> +       [I915_ENGINE_CLASS_RENDER] = "0",
> >> +       [I915_ENGINE_CLASS_COPY] = "1",
> >> +       [I915_ENGINE_CLASS_VIDEO] = "2",
> >> +       [I915_ENGINE_CLASS_VIDEO_ENHANCE] = "3",
> >> +};
> > 
> > Hmm. /sys/class/drm/card0/clients/0/busy/0
> > 
> > Ok. I was worried this was 0/0 and so very bland and liable to clash
> > later.
> > 
> >> +
> >>   int
> >>   __i915_drm_client_register(struct i915_drm_client *client,
> >>                             struct task_struct *task)
> >>   {
> >>          struct i915_drm_clients *clients = client->clients;
> >> +       struct drm_i915_private *i915 =
> >> +               container_of(clients, typeof(*i915), clients);
> >> +       struct intel_engine_cs *engine;
> >>          struct device_attribute *attr;
> >> -       int ret = -ENOMEM;
> >> +       int i, ret = -ENOMEM;
> >>          char idstr[32];
> >>   
> >>          if (!clients->root)
> >> @@ -77,10 +130,71 @@ __i915_drm_client_register(struct i915_drm_client *client,
> >>          if (ret)
> >>                  goto err_attr;
> >>   
> >> +       if (HAS_LOGICAL_RING_CONTEXTS(i915)) {
> >> +               client->busy_root =
> >> +                       kobject_create_and_add("busy", client->root);
> >> +               if (!client->busy_root)
> >> +                       goto err_attr;
> >> +
> >> +               for (i = 0; i < ARRAY_SIZE(uabi_class_names); i++) {
> >> +                       struct i915_engine_busy_attribute *i915_attr =
> >> +                               &client->attr.busy[i];
> > 
> > 
> > if (!intel_engine_lookup_user(i915, i, 0))
> >       continue;
> > 
> > i.e. skip if we don't have any engines of that class in the system.
> 
> Yes, thanks.
> 
> >> +
> >> +                       i915_attr->client = client;
> >> +                       i915_attr->engine_class = i;
> >> +
> >> +                       attr = &i915_attr->attr;
> >> +
> >> +                       sysfs_attr_init(&attr->attr);
> >> +
> >> +                       attr->attr.name = uabi_class_names[i];
> >> +                       attr->attr.mode = 0444;
> >> +                       attr->show = show_client_busy;
> >> +
> >> +                       ret = sysfs_create_file(client->busy_root,
> >> +                                               (struct attribute *)attr);
> >> +                       if (ret)
> >> +                               goto err_busy;
> >> +               }
> >> +
> >> +               /* Enable busy stats on all engines. */
> >> +               i = 0;
> >> +               for_each_uabi_engine(engine, i915) {
> >> +                       ret = intel_enable_engine_stats(engine);
> > 
> > Hmm. We gave it a global bit in
> > 
> >       i915->caps.scheduler & I915_SCHEDULER_CAP_ENABLED.
> > 
> > That'll avoid having to do the individual checking and rollback.
> 
> I could add a top level check as a short circuit, but I prefer to check 
> return code from intel_enable_engine_stats since it returns one.

My suggestion was to remove the return code and make it bug out, as we
[can] check before use in i915_pmu.c as well.

> Also if new GuC will have I915_SCHEDULER_CAP_ENABLED it will still fail 
> to enable engine stats and then fallback to pphwsp has to happen.

Brainfart, CAP_SUPPORTS_STATS. 
-Chris