Re: [PATCH] arm64: mm: only initialize swiotlb when necessary

[Catalin Marinas]

On Tue, Jun 21, 2016 at 10:13:45PM +0200, Arnd Bergmann wrote:
> On Tuesday, June 21, 2016 5:06:25 PM CEST Catalin Marinas wrote:
> > On Wed, Jun 08, 2016 at 05:49:59PM +0200, Arnd Bergmann wrote:
> > > On Wednesday, June 8, 2016 1:08:29 PM CEST Catalin Marinas wrote:
> > > > On Wed, Jun 08, 2016 at 03:53:46PM +0800, Jisheng Zhang wrote:
> > > > >  static int __init arm64_dma_init(void)
> > > > >  {
> > > > > +     if (swiotlb_force || max_pfn > (arm64_dma_phys_limit >> PAGE_SHIFT))
> > > > > +             swiotlb = 1;
> > > > > +
> > > > >       return atomic_pool_init();
> > > > >  }
> > > > 
> > > > So any platform with RAM larger than 4GB would still initialise swiotlb.
> > > > I wouldn't say it's an issue, 64MB is not a significant loss on such
> > > > systems.
> > > > 
> > > > An alternative would be to defer the freeing until we are aware of all
> > > > possible dma masks for the populated devices (e.g. from DT), though I'm
> > > > not sure that's enough, drivers may try to change such masks when
> > > > probed.
> > > 
> > > Right, this is a hard problem, because you can in theory have odd devices
> > > that require a DMA mask lower than the limit of ZONE_DMA.
> > 
> > I'm not sure what we can do about such devices even with swiotlb. The
> > bounce buffer is allocated from ZONE_DMA and currently it assumes a
> > 32-bit mask from the start of RAM, so it is not guaranteed to support a
> > smaller mask. We may need to come up with some DT memory node attribute
> > to define the minimum DMA-able memory and restrict ZONE_DMA during early
> > boot but I would rather wait until we hit a real issue in practice.
> 
> The bounce buffer is allocated at early boot time in order to get an address
> as low as possible, in the hope that it is small enough for those cases.

The swiotlb allocation is triggered from mem_init() and that's well
after the memblock has been initialised. The swiotlb_init() function
uses memblock_virt_alloc_low_nopanic() which is a top-down allocator
with an upper bound of ARCH_LOW_ADDRESS_LIMIT. The latter is set by the
arm64 code to the top 32-bit of RAM (plus some offset), so it's likely
that the swiotlb bounce buffer will be placed in the upper range of the
32-bit mask.

-- 
Catalin