On Thu, Nov 22, 2018 at 08:50:47AM -0800, Linus Torvalds wrote:

On Thu, Nov 22, 2018 at 6:03 AM Christoph Hellwig [off-list ref] wrote:

quoted

The 0 return doesn't work for direct mappings that have ram at address
zero and a lot of IOMMUs that start allocating bus space from address
zero, so we can't consolidate on that, but I think we can move everyone
to all-Fs, which the patch here does.

Hmm. Maybe not limit it to just all ones, but actually use the
(standard, for the kernel) IS_ERR_VALUE()?

That basically reserves the last 4095 values in an unsigned long for
error values.

Then those functions could actually return *what* error they
encountered, using just plain

        return -ENOMEM;

or whatever?

Linus,

I'm afraid that won't work very well - 32 bit platforms with 64-bit
addresses (LPAE) would have dma_addr_t as a 64-bit value, which
wouldn't fit into an unsigned long.

IS_ERR_VALUE() would cast a 64-bit DMA address down to a 32-bit
pointer (effectively masking with 0xffffffff).  It would have the
effect of making (eg) 0xXXXXXXXX_fffffVVV an error, where XXXXXXXX
are any of the top 32-bits of a 64-bit bus address, and VVV is the
error code value.

That could be a problem if you hit it in several places throughout
your available RAM... we'd have to mark every top page of RAM in
a naturally aligned 4GB as unusable, as well as block the top page
in natually aligned 4GB blocks from IOMMUs... and then what about
buses that have weird offsets...

So, I don't think the IS_ERR_VALUE() would work very well.

-- 
RMK's Patch system: http://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
According to speedtest.net: 11.9Mbps down 500kbps up