1MB alignment dma_alloc question

1MB alignment dma_alloc question

[Russell McGuire]

Anyone,

Putting together a driver, IMA for ATM in a MPC83xx. Anyway the manual
states that one of the data structures for the QE needs to be allocated on a
1MB boundary in external memory.

Normally these come in 128/64/32byte boundaries and are easy to self align,
with some bit logic. However, after reading up on dma_alloc_coherent I see
that a limitation is that dma cannot span 64K page boundaries.

So this begs a question is their a better way to get a 1MB aligned table?
Other than allocating almost 2MB of memory and then trying to mask off the
LSBs? Seems like an awful waste of memory. I guess the real question is
there a better kernel call that returns already aligned DMA memory?

Thanks,

-Russ

Re: 1MB alignment dma_alloc question

[Scott Wood]

Russell McGuire wrote:
> Putting together a driver, IMA for ATM in a MPC83xx. Anyway the manual
> states that one of the data structures for the QE needs to be allocated on a
> 1MB boundary in external memory.
> 
> Normally these come in 128/64/32byte boundaries and are easy to self align,
> with some bit logic. However, after reading up on dma_alloc_coherent I see
> that a limitation is that dma cannot span 64K page boundaries.

That sounds like something related to legacy PC DMA.  I doubt there's 
any such limitation in the API itself.

> So this begs a question is their a better way to get a 1MB aligned table?
> Other than allocating almost 2MB of memory and then trying to mask off the
> LSBs? Seems like an awful waste of memory. I guess the real question is
> there a better kernel call that returns already aligned DMA memory?

It looks like allocations will be naturally aligned on powerpc (due to 
coming straight from alloc_pages), though it's probably not guaranteed 
by the API.

-Scott

Re: 1MB alignment dma_alloc question

[Arnd Bergmann]

On Friday 05 September 2008, Russell McGuire wrote:

> So this begs a question is their a better way to get a 1MB aligned table?
> Other than allocating almost 2MB of memory and then trying to mask off the
> LSBs? Seems like an awful waste of memory. I guess the real question is
> there a better kernel call that returns already aligned DMA memory?

__get_free_pages() will allocate naturally aligned pages from the buddy
allocator, so that should do just what you need. If you don't do it at
boot time though,  you might not be able to get a full unfragmented
megabyte any more.

Another option would be to call alloc_bootmem() first, align the buffer
and call free_bootmem() on the areas that you do not need. This has to
be done *very* early in the boot process.

	Arnd <><