Re: Regression due to d98849aff879 (dma-direct: handle DMA_ATTR_NO_KERNEL_MAPPING in common code)

[Lucas Stach]

Am Dienstag, den 06.08.2019, 13:33 +0200 schrieb Christoph Hellwig:
> On Tue, Aug 06, 2019 at 11:13:29AM +0200, Lucas Stach wrote:
> > Hi Christoph,
> > 
> > I just found a regression where my NVMe device is no longer able to
> > set
> > up its HMB.
> > 
> > After subject commit dma_direct_alloc_pages() is no longer
> > initializing
> > dma_handle properly when DMA_ATTR_NO_KERNEL_MAPPING is set, as the
> > function is now returning too early.
> > 
> > Now this could easily be fixed by adding the phy_to_dma translation
> > to
> > the NO_KERNEL_MAPPING code path, but I'm not sure how this stuff
> > interacts with the memory encryption stuff set up later in the
> > function, so I guess this should be looked at by someone with more
> > experience with this code than me.
> 
> There is not much we can do about the memory encryption case here,

Which I would guess means we need to ignore DMA_ATTR_NO_KERNEL_MAPPING
in that case instead of dropping out early?

> as that requires a kernel address to mark the memory as unencrypted.
> 
> So the obvious trivial fix is probably the right one:
> 
> 
> diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
> index 59bdceea3737..c49120193309 100644
> --- a/kernel/dma/direct.c
> +++ b/kernel/dma/direct.c
> @@ -135,6 +135,7 @@ void *dma_direct_alloc_pages(struct device *dev,
> size_t size,
> Â		if (!PageHighMem(page))
> Â			arch_dma_prep_coherent(page, size);
> Â		/* return the page pointer as the opaque cookie */
> +		*dma_handle = phys_to_dma(dev, page_to_phys(page));
> Â		return page;
> Â	}
> Â