[RFC PATCH v3 1/4] RDMA/umem: Support importing dma-buf as user memory region

[Daniel Vetter]

On Mon, Oct 05, 2020 at 04:18:11PM +0000, Xiong, Jianxin wrote:
> > -----Original Message-----
> > From: Jason Gunthorpe <jgg at ziepe.ca>
> > Sent: Monday, October 05, 2020 6:13 AM
> > To: Xiong, Jianxin <jianxin.xiong at intel.com>
> > Cc: linux-rdma at vger.kernel.org; dri-devel at lists.freedesktop.org; Doug Ledford <dledford at redhat.com>; Leon Romanovsky
> > <leon at kernel.org>; Sumit Semwal <sumit.semwal at linaro.org>; Christian Koenig <christian.koenig at amd.com>; Vetter, Daniel
> > <daniel.vetter at intel.com>
> > Subject: Re: [RFC PATCH v3 1/4] RDMA/umem: Support importing dma-buf as user memory region
> > 
> > On Sun, Oct 04, 2020 at 12:12:28PM -0700, Jianxin Xiong wrote:
> > > Dma-buf is a standard cross-driver buffer sharing mechanism that can
> > > be used to support peer-to-peer access from RDMA devices.
> > >
> > > Device memory exported via dma-buf is associated with a file descriptor.
> > > This is passed to the user space as a property associated with the
> > > buffer allocation. When the buffer is registered as a memory region,
> > > the file descriptor is passed to the RDMA driver along with other
> > > parameters.
> > >
> > > Implement the common code for importing dma-buf object and mapping
> > > dma-buf pages.
> > >
> > > Signed-off-by: Jianxin Xiong <jianxin.xiong at intel.com>
> > > Reviewed-by: Sean Hefty <sean.hefty at intel.com>
> > > Acked-by: Michael J. Ruhl <michael.j.ruhl at intel.com>
> > > ---
> > >  drivers/infiniband/core/Makefile      |   2 +-
> > >  drivers/infiniband/core/umem.c        |   4 +
> > >  drivers/infiniband/core/umem_dmabuf.c | 291
> > > ++++++++++++++++++++++++++++++++++
> > >  drivers/infiniband/core/umem_dmabuf.h |  14 ++
> > >  drivers/infiniband/core/umem_odp.c    |  12 ++
> > >  include/rdma/ib_umem.h                |  19 ++-
> > >  6 files changed, 340 insertions(+), 2 deletions(-)  create mode
> > > 100644 drivers/infiniband/core/umem_dmabuf.c
> > >  create mode 100644 drivers/infiniband/core/umem_dmabuf.h
> > 
> > I think this is using ODP too literally, dmabuf isn't going to need fine grained page faults, and I'm not sure this locking scheme is OK - ODP is
> > horrifically complicated.
> > 
> 
> > If this is the approach then I think we should make dmabuf its own stand alone API, reg_user_mr_dmabuf()
> 
> That's the original approach in the first version. We can go back there.
> 
> > 
> > The implementation in mlx5 will be much more understandable, it would just do dma_buf_dynamic_attach() and program the XLT exactly
> > the same as a normal umem.
> > 
> > The move_notify() simply zap's the XLT and triggers a work to reload it after the move. Locking is provided by the dma_resv_lock. Only a
> > small disruption to the page fault handler is needed.
> > 
> 
> We considered such scheme but didn't go that way due to the lack of notification when the move is done and thus the work wouldn't know when it can reload.
> 
> Now I think it again, we could probably signal the reload in the page fault handler. 
> 
> > > +	dma_resv_lock(umem_dmabuf->attach->dmabuf->resv, NULL);
> > > +	sgt = dma_buf_map_attachment(umem_dmabuf->attach,
> > > +				     DMA_BIDIRECTIONAL);
> > > +	dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv);
> > 
> > This doesn't look right, this lock has to be held up until the HW is programmed
> 
> The mapping remains valid until being invalidated again. There is a sequence number checking before programming the HW. 
> 
> > 
> > The use of atomic looks probably wrong as well.
> 
> Do you mean umem_dmabuf->notifier_seq? Could you elaborate the concern?
> 
> > 
> > > +	k = 0;
> > > +	total_pages = ib_umem_odp_num_pages(umem_odp);
> > > +	for_each_sg(umem->sg_head.sgl, sg, umem->sg_head.nents, j) {
> > > +		addr = sg_dma_address(sg);
> > > +		pages = sg_dma_len(sg) >> page_shift;
> > > +		while (pages > 0 && k < total_pages) {
> > > +			umem_odp->dma_list[k++] = addr | access_mask;
> > > +			umem_odp->npages++;
> > > +			addr += page_size;
> > > +			pages--;
> > 
> > This isn't fragmenting the sg into a page list properly, won't work for unaligned things
> 
> I thought the addresses are aligned, but will add explicit alignment here.

Everyone's working under the assumption that dma_buf sg lists are fully
PAGE aligned. Lots of stuff can break otherwise all over the place. You
might get more than that, especially for p2p since many gpus work on 64kb
pages for the vram, but system pages are 4k aligned.

I think we need a patch to clarify this in the kerneldoc, bot for
dma_buf_map_attachment for importers and the dma_buf_ops callback for
exporters.

Also I think it'd be good to verify that when dma api debugging is
enabled, i.e. when dma_buf_map_attachment succeeds, walk the entire sg
list and validate that all dma_addr_t segments are page aligned. And if
not, WARN loudly and fail.
-Daniel

 
> > 
> > And really we don't need the dma_list for this case, with a fixed whole mapping DMA SGL a normal umem sgl is OK and the normal umem
> > XLT programming in mlx5 is fine.
> 
> The dma_list is used by both "polulate_mtt()" and "mlx5_ib_invalidate_range", which are used for XLT programming and invalidating (zapping), respectively.
> 
> > 
> > Jason
> _______________________________________________
> dri-devel mailing list
> dri-devel at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch