Re: [RFC 04/11] dma-mapping: add dma_alloc_attrs_node() and dma_alloc_coherent_node()

["Björn Töpel" <bjorn@kernel.org>]

Jakub Kicinski <kuba@kernel.org> writes:

> Coherent DMA memory is always allocated on the NUMA node of the device.
> That's the right choice for memory mostly accessed by the device, but
> some coherent allocations are accessed by the CPU on every operation.
> For example a NIC's descriptor and completion rings are written and
> read by the CPU servicing the queue, which may well be on a different
> node than the device. Placing such rings on the device's node means
> the CPU accesses remote memory in the fast path.
>
> Drivers work around this today by temporarily changing the node
> of the device with set_dev_node() around dma_alloc_coherent()
> (mlx4, mlx5, ixgbe, xgbe). This is racy against anything else
> looking at the device's node at the same time.
>
> Add dma_alloc_attrs_node() which takes the node explicitly, and
> a dma_alloc_coherent_node() wrapper. dma_alloc_attrs() now passes
> dev_to_node(). The node is only a preference, the same as for
> kmalloc_node(). It is honored by dma-direct and iommu-dma, for both
> the page allocator and the per-numa CMA areas. Allocations from fixed
> pools and from dma_map_ops->alloc implementations ignore it.
>
> The memory is freed with the existing dma_free_attrs() /
> dma_free_coherent(), freeing does not depend on the node.
>
> Signed-off-by: Jakub Kicinski <kuba@kernel.org>
> ---
>  Documentation/core-api/dma-api.rst | 15 ++++++++++++++
>  include/linux/dma-mapping.h        | 14 +++++++++++++
>  kernel/dma/mapping.c               | 32 ++++++++++++++++++++++++++----
>  3 files changed, 57 insertions(+), 4 deletions(-)
>
> diff --git a/Documentation/core-api/dma-api.rst b/Documentation/core-api/dma-api.rst
> index ba23a472f794..f9a32b8c338f 100644
> --- a/Documentation/core-api/dma-api.rst
> +++ b/Documentation/core-api/dma-api.rst
> @@ -57,6 +57,21 @@ The flag parameter allows the caller to specify the ``GFP_`` flags (see
>  kmalloc()) for the allocation (the implementation may ignore flags that affect
>  the location of the returned memory, like GFP_DMA).
>  
> +::
> +
> +	void *
> +	dma_alloc_coherent_node(struct device *dev, size_t size,
> +				dma_addr_t *dma_handle, gfp_t flag, int node)
> +
> +Same as dma_alloc_coherent(), but places the memory on the given NUMA node
> +rather than on the node of the device. This is useful for memory accessed
> +frequently by both the device and a CPU on a different node, for example
> +descriptor rings serviced by that CPU. Pass NUMA_NO_NODE for no preference.

nit: maybe mention that callers that want the device-node semantics need
to pass dev_to_node(dev) themselves?


Björn