[BUG] XHCI getting ZONE_DMA32 memory > than its bus_dma_limit

[Christoph Hellwig]

On Mon, Jul 06, 2020 at 04:09:36PM +0200, Nicolas Saenz Julienne wrote:
> On Sun, 2020-07-05 at 16:41 -0700, David Rientjes wrote:
> > On Fri, 3 Jul 2020, Robin Murphy wrote:
> > > Or perhaps just get rid of atomic_pool_dma32 (and allocate atomic_pool_dma
> > > from ZONE_DMA32 if !ZONE_DMA). That should make it fall pretty much back in
> > > line while still preserving the potential benefit of the kernel pool for
> > > non-address-constrained devices.
> > > 
> > 
> > I assume it depends on how often we have devices where 
> > __dma_direct_alloc_pages() behavior is required, i.e. what requires the 
> > dma_coherent_ok() checks and altering of the gfp flags to get memory that 
> > works.
> > 
> > Is the idea that getting rid of atomic_pool_dma32 would use GFP_KERNEL 
> > (and atomic_pool_kernel) as the default policy here?  That doesn't do any 
> > dma_coherent_ok() checks so dma_direct_alloc_pages would return from 
> > ZONE_NORMAL without a < 3G check?
> 
> IIUC this is not what Robin proposes.
> 
> The idea is to only have one DMA pool, located in ZONE_DMA, if enabled, and
> ZONE_DMA32 otherwise. This way you're always sure the memory is going to be
> good enough for any device while maintaining the benefits of
> atomic_pool_kernel.

That is how I understood the proposal from Robin and I think it is
the right thing to do.

> > It *seems* like we want to check if dma_coherent_ok() succeeds for ret in 
> > dma_direct_alloc_pages() when allocating from the atomic pool and, based 
> > on criteria that allows fallback, just fall into 
> > __dma_direct_alloc_pages()?
> 
> I suspect I don't have enough perspective here but, isn't that breaking the
> point of having an atomic pool? Wouldn't that generate big latency spikes? I
> can see how audio transfers over USB could be affected by this specifically,
> IIRC those are allocated atomically and have timing constraints.
> 
> That said, if Robin solution works for you, I don't mind having a go at it.

We can't just fall back to __dma_direct_alloc_pages when allocation
from the atomic pool fails, as the atomic pool exists for provide
allocations that require sleeping actions for callers that can't
sleep.