[PATCH] dma_declare_coherent_memory: push ioremap() up to caller

[Russell King - ARM Linux]

On Fri, Dec 24, 2010 at 02:55:25PM +0100, Janusz Krzysztofik wrote:
> Friday 24 December 2010 14:02:00 Russell King - ARM Linux wrote:
> > On Fri, Dec 24, 2010 at 12:20:32AM +0100, Janusz Krzysztofik wrote:
> > > The patch tries to implement a solution suggested by Russell King,
> > > http://lists.infradead.org/pipermail/linux-arm-kernel/2010-December
> > >/035264.html. It is expected to solve video buffer allocation issues
> > > for at least a few soc_camera I/O memory less host interface
> > > drivers, designed around the videobuf_dma_contig layer, which
> > > allocates video buffers using dma_alloc_coherent().
> > >
> > > Created against linux-2.6.37-rc5.
> > >
> > > Tested on ARM OMAP1 based Amstrad Delta with a WIP OMAP1 camera
> > > patch, patterned upon two mach-mx3 machine types which already try
> > > to use the dma_declare_coherent_memory() method for reserving a
> > > region of system RAM preallocated with another
> > > dma_alloc_coherent(). Compile tested for all modified files except
> > > arch/sh/drivers/pci/fixups-dreamcast.c.
> >
> > Another note: with the pair of patches I've sent to the
> > linux-arm-kernel list earlier today changing the DMA coherent
> > allocator to steal memory from the system at boot.
> >
> > This means there's less need to pre-allocate DMA memory - if there's
> > sufficient contiguous space in the DMA region to satisfy the
> > allocation, then the allocation will succeed.  It's also independent
> > of the maximum page size from the kernel's memory allocators too.
> >
> > So I suspect that mach-mx3 (and others) no longer need to do their
> > own pre-allocation anymore if both of these patches go in.
> 
> Then, my rationale will no longer be valid. So, either drop my patch if 
> you think you have finally come out with a better solution which 
> doesn't touch any system-wide API, or suggest a new justification for 
> use in the commit log if you think still the patch solves something 
> important.

No.  It's not clear whether my pair of patches are both going to make it
into the kernel, or even what timeframe they're going to make it in.

Irrespective of that, we do need a solution to this problem so that this
stuff starts working again.

In any case, your patch makes complete sense through and through:

1. if other architecture/machine wants to reserve a chunk of DMA-able
   memory for a specific device (eg, because of a restriction on the
   number of DMA address bits available) and it's completely DMA
   coherent, it shouldn't have to go through the pains of having that
   memory unconditionally ioremap'd.

2. What if the memory being provided from some source where you already
   have the mapping setup in a specific way for a reason?

For example, say I have an ARM design, and all peripherals are in a
single 256MB memory region, which includes a chunk of SRAM set aside
for DMA purposes.  Let's say I can map that really efficiently by a
few page table entries, which means relatively little TLB usage for
accessing this region.

With the current API, it becomes difficult to pass that mapping through
the dma_declare_coherent_memory() because the physical address goes
through ioremap(), which obfuscates what's going on.  However, if I
could pass the bus and virtual address, then there's no ambiguity over
what I want to do.

What if I want to declare memory to the DMA coherent allocator with
attributes different from what ioremap() gives me, maybe with write
combining properties (because it happens to be safe for the specific
device) ?

Passing the virtual address allows the API to become much more flexible.
Not only that, it allows it to be used on ARM, rather than becoming (as
it currently stands) prohibited on ARM.

I believe that putting ioremap() inside this API was the wrong thing to
do, and moving it outside makes the API much more flexible and usable.
It's something I still fully support.