[PATCH v10 4/6] RDMA/mlx5: Support dma-buf based userspace memory region

[Jason Gunthorpe]

On Fri, Nov 13, 2020 at 03:51:20AM +0000, Xiong, Jianxin wrote:

> > > +static void mlx5_ib_dmabuf_invalidate_cb(struct dma_buf_attachment
> > > +*attach) {
> > > +	struct ib_umem_dmabuf *umem_dmabuf = attach->importer_priv;
> > > +	struct mlx5_ib_mr *mr = umem_dmabuf->private;
> > > +
> > > +	dma_resv_assert_held(umem_dmabuf->attach->dmabuf->resv);
> > > +
> > > +	if (mr)
> > 
> > I don't think this 'if (mr)' test is needed anymore? I certainly prefer it gone as it is kind of messy. I expect unmapping the dma to ensure this
> > function is not running, and won't run again.
> 
> It is still needed. When the dma-buf moves, the callback function of every attached importer is invoked, regardless if the importer has mapped the dma or not.
> 
> > 
> > > +/**
> > > + * mlx5_ib_fence_dmabuf_mr - Stop all access to the dmabuf MR
> > > + * @mr: to fence
> > > + *
> > > + * On return no parallel threads will be touching this MR and no DMA
> > > +will be
> > > + * active.
> > > + */
> > > +void mlx5_ib_fence_dmabuf_mr(struct mlx5_ib_mr *mr) {
> > > +	struct ib_umem_dmabuf *umem_dmabuf = to_ib_umem_dmabuf(mr->umem);
> > > +
> > > +	/* Prevent new page faults and prefetch requests from succeeding */
> > > +	xa_erase(&mr->dev->odp_mkeys, mlx5_base_mkey(mr->mmkey.key));
> > > +
> > > +	/* Wait for all running page-fault handlers to finish. */
> > > +	synchronize_srcu(&mr->dev->odp_srcu);
> > > +
> > > +	wait_event(mr->q_deferred_work,
> > > +!atomic_read(&mr->num_deferred_work));
> > > +
> > > +	dma_resv_lock(umem_dmabuf->attach->dmabuf->resv, NULL);
> > > +	mlx5_mr_cache_invalidate(mr);
> > > +	umem_dmabuf->private = NULL;
> > > +	ib_umem_dmabuf_unmap_pages(umem_dmabuf);
> > > +	dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv);
> > > +
> > > +	if (!mr->cache_ent) {
> > > +		mlx5_core_destroy_mkey(mr->dev->mdev, &mr->mmkey);
> > > +		WARN_ON(mr->descs);
> > > +	}
> > 
> > I didn't check carefully, but are you sure this destroy_mkey should be here??
> 
> To my understanding, yes. This is similar to what dma_fence_odp_mr() does,
> just inlined here since it's not called from other places.

I think you should put the calls to dma_buf_dynamic_attach() and
dma_buf_detach() into mlx5, it makes the whole thing a little cleaner,
then the umem->private isn't needed any more either.

Jason