Re: [PATCH 19/28] dma-mapping: replace DMA_ATTR_NON_CONSISTENT with dma_{alloc, free}_pages

[Tomasz Figa]

Hi Christoph,

On Wed, Aug 19, 2020 at 8:57 AM Christoph Hellwig <hch@xxxxxx> wrote:
>
> Add a new API to allocate and free pages that are guaranteed to be
> addressable by a device, but otherwise behave like pages allocated by
> alloc_pages.  The intended APIs to sync them for use with the device
> and cpu are dma_sync_single_for_{device,cpu} that are also used for
> streaming mappings.
>
> Switch all drivers over to this new API, but keep the usage of the
> crufty dma_cache_sync API for now, which will be cleaned up on a driver
> by driver basis.
>
> Signed-off-by: Christoph Hellwig <hch@xxxxxx>
> ---
>  Documentation/core-api/dma-api.rst        | 68 +++++++++++------------
>  Documentation/core-api/dma-attributes.rst |  8 ---
>  arch/alpha/kernel/pci_iommu.c             |  2 +
>  arch/arm/mm/dma-mapping-nommu.c           |  2 +
>  arch/arm/mm/dma-mapping.c                 |  4 ++
>  arch/ia64/hp/common/sba_iommu.c           |  2 +
>  arch/mips/jazz/jazzdma.c                  |  7 +--
>  arch/powerpc/kernel/dma-iommu.c           |  2 +
>  arch/powerpc/platforms/ps3/system-bus.c   |  4 ++
>  arch/powerpc/platforms/pseries/vio.c      |  2 +
>  arch/s390/pci/pci_dma.c                   |  2 +
>  arch/x86/kernel/amd_gart_64.c             |  2 +
>  drivers/iommu/dma-iommu.c                 |  2 +
>  drivers/iommu/intel/iommu.c               |  4 ++
>  drivers/net/ethernet/i825xx/lasi_82596.c  | 13 ++---
>  drivers/net/ethernet/seeq/sgiseeq.c       | 12 ++--
>  drivers/parisc/ccio-dma.c                 |  2 +
>  drivers/parisc/sba_iommu.c                |  2 +
>  drivers/scsi/53c700.c                     |  8 +--
>  drivers/scsi/sgiwd93.c                    | 12 ++--
>  drivers/xen/swiotlb-xen.c                 |  2 +
>  include/linux/dma-direct.h                |  5 ++
>  include/linux/dma-mapping.h               | 29 ++++++++--
>  include/linux/dma-noncoherent.h           |  3 -
>  kernel/dma/direct.c                       | 51 ++++++++++++++++-
>  kernel/dma/mapping.c                      | 43 +++++++++++++-
>  kernel/dma/ops_helpers.c                  | 35 ++++++++++++
>  kernel/dma/virt.c                         |  2 +
>  sound/mips/hal2.c                         | 20 +++----
>  29 files changed, 254 insertions(+), 96 deletions(-)
>

Thanks for the patch. The general design looks quite nice, but please
see my comments inline.


> diff --git a/Documentation/core-api/dma-api.rst b/Documentation/core-api/dma-api.rst
> index 90239348b30f6f..047fcfffa0e5cf 100644
> --- a/Documentation/core-api/dma-api.rst
> +++ b/Documentation/core-api/dma-api.rst
> @@ -516,48 +516,53 @@ routines, e.g.:::
>         }
>
>
> -Part II - Advanced dma usage
> -----------------------------
> +Part II - Non-coherent DMA allocations
> +--------------------------------------
>
> -Warning: These pieces of the DMA API should not be used in the
> -majority of cases, since they cater for unlikely corner cases that
> -don't belong in usual drivers.
> +These APIs allow to allocate pages that can be used like normal pages
> +in the kernel direct mapping, but are guaranteed to be DMA addressable.

Could we elaborate a bit more on what "like normal pages in kernel
direct mapping" mean from the driver perspective?

>
>  If you don't understand how cache line coherency works between a
>  processor and an I/O device, you should not be using this part of the
> -API at all.
> +API.
>
>  ::
>
>         void *
> -       dma_alloc_attrs(struct device *dev, size_t size, dma_addr_t *dma_handle,
> -                       gfp_t flag, unsigned long attrs)
> +       dma_alloc_pages(struct device *dev, size_t size, dma_addr_t *dma_handle,
> +                       enum dma_data_direction dir, gfp_t gfp)
> +
> +This routine allocates a region of <size> bytes of consistent memory.  It
> +returns a pointer to the allocated region (in the processor's virtual address
> +space) or NULL if the allocation failed. The returned memory is guanteed to
> +behave like memory allocated using alloc_pages.

There is one aspect that the existing dma_alloc_attrs() handles, but
this new function doesn't: IOMMU support. The function will always
allocate a physically-contiguous block memory, which is a costly
operation and not even guaranteed to succeed, even if enough free
memory is available.

Modern SoCs employ IOMMUs to avoid the need to allocate
physically-contiguous memory and those happen to be also the devices
that could benefit from non-coherent allocations a lot. One of the
tasks of the DMA API was making it possible to allocate suitable
memory for a given device, without having the driver know about the
SoC integration details, such as the presence of an IOMMU.

Today, dma_alloc_attrs() uses the .alloc callback of the dma_ops
struct and the IOMMU-aware implementations, like the dma-iommu helpers
[1], would allocate discontiguous pages. Therefore, while I see the
DMA-aware page allocation functionality as a useful functionality on
its own for scatter-gather-capable hardware, I believe it is not a
complete replacement for dma_alloc_attrs() with the
DMA_ATTR_NON_CONSISTENT attribute.

[1] https://elixir.bootlin.com/linux/v5.9-rc1/source/drivers/iommu/dma-iommu.c#L510

Best regards,
Tomasz