Re: [PATCH v8 0/9] pci: add pci_iomap_wc() and pci_ioremap_wc_bar()

["Luis R. Rodriguez" <mcgrof@suse.com>]

On Fri, Jun 26, 2015 at 12:12:06PM +1000, Benjamin Herrenschmidt wrote:
> On Wed, 2015-06-24 at 18:22 -0700, Luis R. Rodriguez wrote:
> > Although I had test compiled this before just to be safe I went ahead and
> > successfully test-compiled this set with allmodconfig, specially since I've now
> > removed the exports for the devres routines.  Please let me know if these might
> > be able to go through you or if there are any questions. I will note the recent
> > discussion with Benjamin over the v7 series concluded that the ideas we both
> > were alluding to, on automating instead the WC effects for devices seems a bit
> > too idealistic for PCI / PCIE for now, but perhaps we should at least consider
> > this in the future for userspace mmap() calls [4].
> 
> So I've been trying to figure out how to make this practically work for us (powerpc).
> 
> writel() will never write combine for us, it uses too heavy barriers.
> 
> writel_relaxed() today is identical to writel() but we can change it.
> 
> The problem is that switching to G=0 mappings (which is what provides us with write
> combining) also architecturally enables prefetch and speculative loads... and again
> architecturally (the implementations may differ), kills the effect of the lightweight
> io barrier eieio which we would have to use in readl_relaxed() and writel_relaxed()
> to provide their normal semantics.
> 
> So it boils down to: Can we modify the documentation of readl_relaxed() and writel_relaxed()
> to define them as being even further relaxed when using a "wc" mapping ?
>
> Otherwise, the only way out I see for us on powerpc is to bias massively writel_relaxed()
> against real_relaxed() by putting heavy barriers around the load in the latter so we can
> keep them completely out of the former and still enable wc.

Depends if you semantically then also are implicating its use for the ioremap_wc()
area and if we've ensured we've visited all other possibilities to avoid this. Instead
of replying here though it seems we have a large general ioremap() semantic discussion
ongoing on another thread which is far ahead of this one and more generalized. Mind
following up there, seems the party is there:

http://lkml.kernel.org/r/20150707160703.GR7021@wotan.suse.de

 Luis