Re: [PATCH 3/3] cnic,bnx2,bnx2x: use UIO_MEM_DMA_COHERENT

[Paolo Abeni]

On Mon, 2023-10-02 at 10:59 +0200, Hannes Reinecke wrote:
> On 10/2/23 10:46, Greg Kroah-Hartman wrote:
> > On Mon, Oct 02, 2023 at 12:50:21AM -0700, Jerry Snitselaar wrote:
> > > On Mon, Oct 02, 2023 at 08:04:24AM +0200, Christoph Hellwig wrote:
> > > > On Sun, Oct 01, 2023 at 07:22:36AM -0700, Jerry Snitselaar wrote:
> > > > > Changes last year to the dma-mapping api to no longer allow __GFP_COMP,
> > > > > in particular these two (from the e529d3507a93 dma-mapping pull for
> > > > > 6.2):
> > > > 
> > > > That's complete BS.  The driver was broken since day 1 and always
> > > > ignored the DMA API requirement to never try to grab the page from the
> > > > dma coherent allocation because you generally speaking can't.  It just
> > > > happened to accidentally work the trivial dma coherent allocator that
> > > > is used on x86.
> > > > 
> > > 
> > > re-sending since gmail decided to not send plain text:
> > > 
> > > Yes, I agree that it has been broken and misusing the API. Greg's
> > > question was what changed though, and it was the clean up of
> > > __GFP_COMP in dma-mapping that brought the problem in the driver to
> > > light.
> > > 
> > > I already said the other day that cnic has been doing this for 14
> > > years. I'm not blaming you or your __GFP_COMP cleanup commits, they
> > > just uncovered that cnic was doing something wrong. My apologies if
> > > you took it that way.
> > 
> > As these devices aren't being made anymore, and this api is really not a
> > good idea in the first place, why don't we just leave it broken and see
> > if anyone notices?
> > 
> Guess what triggered this mail thread.
> Some customers did notice.
> 
> Problem is that these devices were built as the network interface in 
> some bladecenter machines, so you can't just replace them with a 
> different Ethernet card.

This route looks a no-go.

Out of sheer ignorance, would the iommu hack hinted in the cover letter
require similar controversial changes?

Cheers,

Paolo