Re: DMA direct mapping fix for 5.4 and earlier stable branches

[Robin Murphy]

On 2021-02-09 12:36, Sumit Garg wrote:
> Hi Christoph,
>
> On Tue, 9 Feb 2021 at 15:06, Christoph Hellwig <hch@xxxxxx> wrote:
> > On Tue, Feb 09, 2021 at 10:23:12AM +0100, Greg KH wrote:
> > > >    From the view point of ZeroCopy using DMABUF, is 5.4 not
> > > > mature enough, and is 5.10 enough mature ?
> > > >    This is the most important point for judging migration.
> > >
> > > How do you judge "mature"?
> > >
> > > And again, if a feature isn't present in a specific kernel version, why
> > > would you think that it would be a viable solution for you to use?
> >
> > I'm pretty sure dma_get_sgtable has been around much longer and was
> > supposed to work, but only really did work properly for arm32, and
> > for platforms with coherent DMA.  I bet he is using non-coherent arm64,
>
> It's an arm64 platform using coherent DMA where device coherent DMA
> memory pool is defined in the DT as follows:
>
>          reserved-memory {
>                  #address-cells = <2>;
>                  #size-cells = <2>;
>                  ranges;
>
>                  <snip>
>                  encbuffer: encbuffer@0xb0000000 {
>                          compatible = "shared-dma-pool";
>                          reg = <0 0xb0000000 0 0x08000000>; // this
> area used with dma-coherent
>                          no-map;
>                  };
>                  <snip>
>          };
>
> Device is dma-coherent as per following DT property:
>
>                  codec {
>                          compatible = "socionext,uniphier-pxs3-codec";
>                          <snip>
>                          memory-region = <&encbuffer>;
>                          dma-coherent;
>                          <snip>
>                  };
>
> And call chain to create device coherent DMA pool is as follows:
>
> rmem_dma_device_init();
>    dma_init_coherent_memory();
>      memremap();
>        ioremap_wc();
>
> which simply maps coherent DMA memory into vmalloc space on arm64.
>
> The thing I am unclear is why this is called a new feature rather than
> a bug in dma_common_get_sgtable() which is failing to handle vmalloc
> addresses? While at the same time DMA debug APIs specifically handle
> vmalloc addresses [1].

It's not a bug, it's a fundamental design failure. dma_get_sgtable() has 
only ever sort-of-worked for DMA buffers that come from CMA or regular 
page allocations. In particular, a "no-map" DMA pool is not backed by 
kernel memory, so does not have any corresponding page structs, so it's 
impossible to generate a *valid* scatterlist to represent memory from 
that pool, regardless of what you might get away with provided you don't 
poke too hard at it.

It is not a good API...

Robin.

> [1] https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/kernel/dma/debug.c?h=linux-5.4.y#n1462
>
> -Sumit
>
> > and it would be broken for other drivers there as well if people did
> > test them, which they apparently so far did not.