Re: [PATCH v3 2/2] dma-pool: Only allocate from CMA when in same memory zone

[Nicolas Saenz Julienne]

On Fri, 2020-08-14 at 08:06 +0200, Christoph Hellwig wrote:
> On Sat, Aug 08, 2020 at 08:33:54AM +0200, Christoph Hellwig wrote:
> > On Fri, Aug 07, 2020 at 10:50:19AM +0200, Nicolas Saenz Julienne wrote:
> > > On Fri, 2020-08-07 at 07:21 +0200, Christoph Hellwig wrote:
> > > > On Thu, Aug 06, 2020 at 08:47:55PM +0200, Nicolas Saenz Julienne wrote:
> > > > > There is no guarantee to CMA's placement, so allocating a zone specific
> > > > > atomic pool from CMA might return memory from a completely different
> > > > > memory zone. To get around this double check CMA's placement before
> > > > > allocating from it.
> > > > 
> > > > As the builtbot pointed out, memblock_start_of_DRAM can't be used from
> > > > non-__init code.  But lookig at it I think throwing that in
> > > > is bogus anyway, as cma_get_base returns a proper physical address
> > > > already.
> > > 
> > > It does indeed, but I'm comparing CMA's base with bitmasks that don't take into
> > > account where the memory starts. Say memory starts at 0x80000000, and CMA falls
> > > into ZONE_DMA [0x80000000 0xC0000000], if you want to compare it with
> > > DMA_BIT_MASK(zone_dma_bits) you're forced to unify the memory bases.
> > > 
> > > That said, I now realize that this doesn't work for ZONE_DMA32 which has a hard
> > > limit on 32bit addresses reglardless of the memory base.
> > > 
> > > That said I still need to call memblock_start_of_DRAM() any suggestions WRT
> > > that? I could save the value in dma_atomic_pool_init(), which is __init code.
> > 
> > The pool must be about a 32-bit physical address.  The offsets can be
> > different for every device..

I now see what you mean.

I was trying to blindly fit CMA with arm64's DMA zone setup, which, as it turns
out, doesn't really honor its purpose. arm64 introduced ZONE_DMA to provide a
30-bit address space, but we're creating it regardless of whether it exists or
not. This creates a mismatch between zone_dma_bits and ZONE_DMA's real
placement. I'll try to look at fixing that in arm64.

> Do you plan to resend this one without the memblock_start_of_DRAM
> thingy?

Yes, sorry for the wait, I've been on vacation and short on time, I'll send it
during the day.

Regards,
Nicolas