On Wed, Jan 31, 2024 at 01:29:56PM -0800, Greg Kroah-Hartman wrote:
> On Wed, Jan 31, 2024 at 11:17:32AM -0800, Chris Leech wrote:
> > Use the UIO_MEM_DMA_COHERENT type to properly handle mmap for
> > dma_alloc_coherent buffers.
> >
> > The cnic l2_ring and l2_buf mmaps have caused page refcount issues as
> > the dma_alloc_coherent no longer provide __GFP_COMP allocation as per
> > commit "dma-mapping: reject __GFP_COMP in dma_alloc_attrs".
> >
> > Fix this by having the uio device use dma_mmap_coherent.
> >
> > The bnx2 and bnx2x status block allocations are also dma_alloc_coherent,
> > and should use dma_mmap_coherent. They don't allocate multiple pages,
> > but this interface does not work correctly with an iommu enabled unless
> > dma_mmap_coherent is used.
> >
> > Fixes: bb73955c0b1d ("cnic: don't pass bogus GFP_ flags to dma_alloc_coherent")
>
> This is really the commit that broke things? By adding this, are you
> expecting anyone to backport this change to older kernels?
Well, the driver has literally been broken since day 1. The above
commit is what made people finally care as it also broke on more
common setups. So I'm not sure the fixes tag is correct.
