Re: [PATCH 2/2] drm/i915: provide module parameter description

[Chris Wilson <chris@chris-wilson.co.uk>]

On Wed,  6 Jul 2011 15:14:53 -0700, Ben Widawsky <ben@bwidawsk.net> wrote:
>  bool i915_enable_hangcheck __read_mostly = true;
>  module_param_named(enable_hangcheck, i915_enable_hangcheck, bool, 0644);
> +MODULE_PARM_DESC(enable_hangcheck, "Periodically check GPU activity for "
> +				   "detecting hangs (default: true)");

I think we need to stress here that disabling the hangcheck can cause the
system to hang (X, VT switching, etc) following an GPU hang. The
assumption that the hangcheck will prevail and break any waits (and
eventually release the struct_mutex) is deeply ingrained into the code.

The alternate scenario is to wait without the mutex, which then requires
revalidating all state after a wait, effectively we would have to restart
the operation from the being. And introduces unfairness (potential
starvation) into command/resource allocations...
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre