On Mon, 2017-03-20 at 10:24 +0000, Will Deacon wrote:

Just to be clear here: I'm not against exposing the proc interface if
something actually needs it, but all the requests we've had for this have
been concerned only with the sysfs API. So I'd rather start with just that,
instead of exposing both and have new software written to the proc interface,
which we certainly want to discourage.

The point is that they are tied together. An architecture provides its
pci_mmap_page_range() function and sets HAVE_PCI_MMAP, and then the
*generic* code in drivers/pci will use it from both places.

Although as noted, there is a third userspace API in drivers/vfio which
does this *without* any arch-specific code (using pgprot_noncached(),
and failing to set up the VMA correctly on HAVE_IOREMAP_PROT platforms.

I was looking at cleaning that whole mess up, but got distracted by
kdump-on-CPU-1 failures. If I'm going to blame that on firmware, I'll
get back to PCI mmap this week... :)
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/x-pkcs7-signature
Size: 4938 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20170320/5f0a1750/attachment.bin>