Re: [PATCH 1/2] mm: export find_extend_vma and handle_mm_fault for driver use

[Jesse Barnes <jbarnes@virtuousgeek.org>]

On Thu, 6 Nov 2014 14:01:22 +0100
Joerg Roedel <jroedel@suse.de> wrote:

> On Wed, Nov 05, 2014 at 01:51:09PM -0800, Jesse Barnes wrote:
> > On Wed, 5 Nov 2014 13:03:51 +0100
> > Joerg Roedel <jroedel@suse.de> wrote:
> > > Thanks for testing Oded. Jesse, the patch looks good to me, except the
> > > task accounting for the page-faults. I'd like to get rid of using
> > > task_struct in the IOMMUv2 driver entirely if possible. Also it is not
> > > really the CPU task causing the faults, but some non-CPU process.
> > 
> > Hm, but the CPU task initiates the activity on the GPU, so we should
> > account for it somewhere, right?  I guess I had been thinking of the
> > "task" as spanning the CPUs and GPUs and other devices in the system,
> > rather than just representing the CPU activity.
> 
> One problem is that the task that called amd_iommu_bind_pasid() isn't
> necessarily the same task (thread) that queues the jobs to the device.
> The thread that called amd_iommu_bind_pasid() might even exit while
> other threads still use the mappings.
> 
> Besides that, from an abstract point of view, what is running on the
> device (GPU) is a logically seperate 'thread' of the process which we
> should account for seperatly. If we want accounting for these off-CPU
> threads we should probably introduce some concept of a non-CPU task to
> the kernel and do the accounting there?

Yeah those are good points; I hadn't been thinking of multi-threaded
stuff.  Logically the GPU stuff really is a separate thread in that
sense, so monitoring faults separately makes sense.

I wonder if we need a new "device_task_struct" or
"coprocessor_task_struct" or something to wrap some simple stuff on
non-CPU devices.  They could be sub-classed by drivers that needed
additional driver specific data.

-- 
Jesse Barnes, Intel Open Source Technology Center