Re: [PATCH 2/5] kernel.h: Add non_block_start/end()

[Jason Gunthorpe]

On Thu, Aug 15, 2019 at 12:32:38PM -0400, Jerome Glisse wrote:
> On Thu, Aug 15, 2019 at 12:10:28PM -0300, Jason Gunthorpe wrote:
> > On Thu, Aug 15, 2019 at 04:43:38PM +0200, Daniel Vetter wrote:
> > 
> > > You have to wait for the gpu to finnish current processing in
> > > invalidate_range_start. Otherwise there's no point to any of this
> > > really. So the wait_event/dma_fence_wait are unavoidable really.
> > 
> > I don't envy your task :|
> > 
> > But, what you describe sure sounds like a 'registration cache' model,
> > not the 'shadow pte' model of coherency.
> > 
> > The key difference is that a regirstationcache is allowed to become
> > incoherent with the VMA's because it holds page pins. It is a
> > programming bug in userspace to change VA mappings via mmap/munmap/etc
> > while the device is working on that VA, but it does not harm system
> > integrity because of the page pin.
> > 
> > The cache ensures that each initiated operation sees a DMA setup that
> > matches the current VA map when the operation is initiated and allows
> > expensive device DMA setups to be re-used.
> > 
> > A 'shadow pte' model (ie hmm) *really* needs device support to
> > directly block DMA access - ie trigger 'device page fault'. ie the
> > invalidate_start should inform the device to enter a fault mode and
> > that is it.  If the device can't do that, then the driver probably
> > shouldn't persue this level of coherency. The driver would quickly get
> > into the messy locking problems like dma_fence_wait from a notifier.
> 
> I think here we do not agree on the hardware requirement. For GPU
> we will always need to be able to wait for some GPU fence from inside
> the notifier callback, there is just no way around that for many of
> the GPUs today (i do not see any indication of that changing).

I didn't say you couldn't wait, I was trying to say that the wait
should only be contigent on the HW itself. Ie you can wait on a GPU
page table lock, and you can wait on a GPU page table flush completion
via IRQ.

What is troubling is to wait till some other thread gets a GPU command
completion and decr's a kref on the DMA buffer - which kinda looks
like what this dma_fence() stuff is all about. A driver like that
would have to be super careful to ensure consistent forward progress
toward dma ref == 0 when the system is under reclaim. 

ie by running it's entire IRQ flow under fs_reclaim locking.

> associated with the mm_struct. In all GPU driver so far it is a short
> lived lock and nothing blocking is done while holding it (it is just
> about updating page table directory really wether it is filling it or
> clearing it).

The main blocking I expect in a shadow PTE flow is waiting for the HW
to complete invalidations of its PTE cache.

> > It is important to identify what model you are going for as defining a
> > 'registration cache' coherence expectation allows the driver to skip
> > blocking in invalidate_range_start. All it does is invalidate the
> > cache so that future operations pick up the new VA mapping.
> > 
> > Intel's HFI RDMA driver uses this model extensively, and I think it is
> > well proven, within some limitations of course.
> > 
> > At least, 'registration cache' is the only use model I know of where
> > it is acceptable to skip invalidate_range_end.
> 
> Here GPU are not in the registration cache model, i know it might looks
> like it because of GUP but GUP was use just because hmm did not exist
> at the time.

It is not because of GUP, it is because of the lack of
invalidate_range_end. A driver cannot correctly implement the SPTE
model without invalidate_range_end, even if it holds the page pins via
GUP.

So, I've been assuming the few drivers without invalidate_range_end
are trying to do registration caching, rather than assuming they are
broken.

Jason