[BUG] XHCI getting ZONE_DMA32 memory > than its bus_dma_limit

[Robin Murphy]

On 2020-07-03 15:53, Nicolas Saenz Julienne wrote:
> Hi Jeremy,
> thanks for the bug report.
> 
> Just for the record the offending commit is: c84dc6e68a1d2 ("dma-pool: add
> additional coherent pools to map to gfp mask").
> 
> On Thu, 2020-07-02 at 12:49 -0500, Jeremy Linton wrote:
>> Hi,
>>
>> Using 5.8rc3:
>>
>> The rpi4 has a 3G dev->bus_dma_limit on its XHCI controller. With a usb3
>> hub, plus a few devices plugged in, randomly devices will fail
>> operations. This appears to because xhci_alloc_container_ctx() is
>> getting buffers > 3G via dma_pool_zalloc().
>>
>> Tracking that down, it seems to be caused by dma_alloc_from_pool() using
>> dev_to_pool()->dma_direct_optimal_gfp_mask() to "optimistically" select
>> the atomic_pool_dma32 but then failing to verify that the allocations in
>> the pool are less than the dev bus_dma_limit.
> 
> I can reproduce this too.
> 
> The way I see it, dev_to_pool() wants a strict dma_direct_optimal_gfp_mask()
> that is never wrong, since it's going to stick to that pool for the device's
> lifetime. I've been looking at how to implement it, and it's not so trivial as
> I can't see a failproof way to make a distinction between who needs DMA32 and
> who is OK with plain KERNEL memory.
> 
> Otherwise, as Jeremy points out, the patch needs to implement allocations with
> an algorithm similar to __dma_direct_alloc_pages()'s, which TBH I don't know if
> it's a little overkill for the atomic context.
> 
> Short of finding a fix in the coming rc's, I suggest we revert this.

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.

Robin.