[PATCH v6 1/4] RDMA/umem: Support importing dma-buf as user memory region

[Jason Gunthorpe]

On Fri, Oct 23, 2020 at 09:39:58AM -0700, Jianxin Xiong wrote:
> +/*
> + * Generate a new dma sg list from a sub range of an existing dma sg list.
> + * Both the input and output have their entries page aligned.
> + */
> +static int ib_umem_dmabuf_sgt_slice(struct sg_table *sgt, u64 offset,
> +				    u64 length, struct sg_table *new_sgt)
> +{
> +	struct scatterlist *sg, *new_sg;
> +	u64 start, end, off, addr, len;
> +	unsigned int new_nents;
> +	int err;
> +	int i;
> +
> +	start = ALIGN_DOWN(offset, PAGE_SIZE);
> +	end = ALIGN(offset + length, PAGE_SIZE);
> +
> +	offset = start;
> +	length = end - start;
> +	new_nents = 0;
> +	for_each_sgtable_dma_sg(sgt, sg, i) {
> +		len = sg_dma_len(sg);
> +		off = min(len, offset);
> +		len -= off;
> +		len = min(len, length);
> +		if (len)
> +			new_nents++;
> +		length -= len;
> +		offset -= off;
> +	}
> +
> +	err = sg_alloc_table(new_sgt, new_nents, GFP_KERNEL);
> +	if (err)
> +		return err;

I would really rather not allocate an entirely new table just to take
a slice of an existing SGT. Ideally the expoter API from DMA buf would
prepare the SGL slice properly instead of always giving a whole
buffer.

Alternatively making some small edit to rdma_umem_for_each_dma_block()
and ib_umem_find_best_pgsz() would let it slice the SGL at runtime

You need to rebase on top of this series:

https://patchwork.kernel.org/project/linux-rdma/list/?series=370437

Which makes mlx5 use those new APIs

Jason