Re: [Linux-c6x-dev] [PATCH 3/9] c6x: Provide dma_mmap_coherent()and dma_get_sgtable()

[James Bottomley]

On Tue, 2013-01-22 at 11:33 +0100, Marek Szyprowski wrote:
> Hello,
> 
> On 1/22/2013 11:13 AM, James Bottomley wrote:
> > On Mon, 2013-01-21 at 22:59 +0000, James Bottomley wrote:
> > > On Mon, 2013-01-21 at 21:00 +0100, Geert Uytterhoeven wrote:
> > > > On Tue, Jan 15, 2013 at 5:56 PM, James Bottomley
> > > > <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx> wrote:
> > > > > On Tue, 2013-01-15 at 15:07 +0100, Marek Szyprowski wrote:
> > > > >> On 1/15/2013 10:13 AM, Geert Uytterhoeven wrote:
> > > > >> > Marek?
> > > > >> >
> > > > >> > On Tue, Jan 15, 2013 at 5:16 AM, Vineet Gupta
> > > > >> > <Vineet.Gupta1@xxxxxxxxxxxx> wrote:
> > > > >> > > On Monday 14 January 2013 09:07 PM, Mark Salter wrote:
> > > > >> > >> On Sun, 2013-01-13 at 11:44 +0100, Geert Uytterhoeven wrote:
> > > > >> > >>> c6x/allmodconfig (assumed):
> > > > >> > >>>
> > > > >> > >>> drivers/media/v4l2-core/videobuf2-dma-contig.c: In function âvb2_dc_mmapâ:
> > > > >> > >>> drivers/media/v4l2-core/videobuf2-dma-contig.c:204: error: implicit declaration of function âdma_mmap_coherentâ
> > > > >> > >>> drivers/media/v4l2-core/videobuf2-dma-contig.c: In function âvb2_dc_get_base_sgtâ:
> > > > >> > >>> drivers/media/v4l2-core/videobuf2-dma-contig.c:387: error: implicit declaration of function âdma_get_sgtableâ
> > > > >> > >>>
> > > > >> > >>> For architectures using dma_map_ops, dma_mmap_coherent() and
> > > > >> > >>> dma_get_sgtable() are provided in <asm-generic/dma-mapping-common.h>.
> > > > >> > >>>
> > > > >> > >>> C6x does not use dma_map_ops, hence it should implement them as inline
> > > > >> > >>> stubs using dma_common_mmap() and dma_common_get_sgtable().
> > > > >> > >>>
> > > > >> > >> So are dma_mmap_coherent() and dma_get_sgtable() part of the DMA API
> > > > >> > >> now? I don't them in Documentation/DMA*.txt anywhere.
> > > > >> > >>
> > > > >> > >> Why does the default dma_common_mmap() for !CONFIG_MMU return an
> > > > >> > >> error?
> > > > >> > >>
> > > > >> > >> Wouldn't it be better to provide default implementations that an arch
> > > > >> > >> could override rather than having to patch all "no dma_map_ops"
> > > > >> > >> architectures?
> > > > >> > >>
> > > > >> > > Speaking for the still-reviewed ARC Port, I completely agree with Mark.
> > > > >>
> > > > >> dma_mmap_coherent() was partially in the DMA mapping API for some time, but
> > > > >> it was available only on a few architectures (afair ARM, powerpc and avr32).
> > > > >> This caused significant problems for writing unified device drivers or some
> > > > >> device helper modules which were aimed to work on more than one
> > > > >> architecture.
> > > > >>
> > > > >> dma_get_sgtable() is an extension discussed during the Linaro meetings. It
> > > > >> is required to correctly implement buffer sharing between device driver
> > > > >> without hacks or any assumptions about memory layout in the device drivers.
> > > > >>
> > > > >> I have implemented some generic code for both of those two functions,
> > > > >> keeping
> > > > >> in mind that on some hardware architectures (like already mentioned VIVT)
> > > > >> it might be not possible to provide coherent mapping to userspace. It is
> > > > >> perfectly fine for those functions to return an error in such case.
> > > > >
> > > > > It's not possible on VIPT either.  This means that the API is unusable
> > > > > on quite a large number of architectures.  Surely, if we're starting to
> > > > > write drivers using this, we need to fix the API before more people try
> > > > > to use it.
> > > > >
> > > > > For PA-RISC (and all other VIPT, I assume) I need an API which allows me
> > > > > to remap the virtual address of the kernel component (probably using the
> > > > > kmap area) so the user space and kernel space addresses are congruent.
> > > >
> > > > So what are we gonna do for 3.8? I'd like to get my allmodconfig build
> > > > green again ;-)
> > > >
> > > > Change the API?
> > >
> > > Well, if we want the API to work universally, we have to.  As I said,
> > > for VIPT systems, the only coherency mechanism we have is the virtual
> > > address ... we have to fix that in the kernel to be congruent with the
> > > userspace virtual address if we want coherency between the kernel and
> > > userspace.
> > >
> > > However, if it only needs to work on ARM and x86, it can stay the way it
> > > is and we could just pull it out of the generic core.
> > >
> > > Who actually wants to use this API, and what for?
> > >
> > > > Keep the API but do a best-effort fix to unbreak the builds?
> > > >   - Apply my patches that got acks (avr32/blackfin/cris/m68k),
> > > >   - Use static inlines that return -EINVAL for the rest
> > > > (c6x/frv/mn10300/parisc/xtensa).
> > > > I still have a few m68k fixes queued for 3.8, for which I've been postponing the
> > > > pull request to get this sorted out, so I could include the above.
> > > >
> > > > Any other solution?
> > >
> > > If it's an API that only works on ARM and x86, there's not much point
> > > pretending it's universal, so we should remove it from the generic arch
> > > code and allow only those architectures which can use it.
> >
> > There might be a simple solution:  just replace void *cpu_addr with void
> > **cpu_addr in the API.  This is a bit nasty since compilers think that
> > void ** to void * conversion is quite legal, so it would be hard to pick
> > up misuse of this (uint8_t ** would be better).  That way VIPT could
> > remap the kernel pages to a coherent address.  This would probably have
> > to change in the dma_mmap_attr() and dma_ops structures.
> 
> Frankly, I don't get this change.
> 
> The call sequence for the current api is following:
> 
> /* driver initialization */
> dma_addr_t dmah;
> void *cpu_ptr = dma_alloc_coherent(dev, size, &dmah, GFP_KERNEL);
> 
> /* mmap implementation */
> 
> return dma_mmap_coherent(dev, vma, cpu_addr, dmah, size);
> 
> /* cleanup */
> dma_free_coherent(dev, size, cpu_addr, dmah);
> 
> I assume you want to let dma_mmap_coherent to change the cpu_addr returned
> by dma_alloc_coherent()? I really don't see how this might help.

That's right.

Because on PA-RISC, I can push the current cpu_addr TLB out of the MMU,
do a remap on the physical page to a kernel virtual address which is
congruent with that page in the VMA and we have automatic kernel to user
space coherency without having to take the performance hit of marking
the page as uncacheable and actually getting it to work on both non-U
and no-MMU architectures.  I think we can even update the offset mapping
to generate a fault if someone then tries to address the page via its
old offset address.

I don't quite get the concern: this is what we do today for vmap and
vmalloc.

James


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/