On Mon, 2015-11-02 at 08:10 +1100, Benjamin Herrenschmidt wrote:
> On Sun, 2015-11-01 at 09:45 +0200, Shamir Rabinovitch wrote:
> > Not sure this use case is possible for Infiniband where application hold
> > the data buffers and there is no way to force application to re use the 
> > buffer as suggested.
> > 
> > This is why I think there will be no easy way to bypass the DMA mapping cost
> > for all use cases unless we allow applications to request bypass/pass through
> > DMA mapping (implicitly or explicitly).
> 
> But but but ...
> 
> What I don't understand is how that brings you any safety.
> 
> Basically, either your bridge has a bypass window, or it doesn't. (Or
> it has one and it's enabled or not, same thing).
> 
> If you request the bypass on a per-mapping basis, you basically have to
> keep the window always enabled, unless you do some nasty refcounting of
> how many people have a bypass mapping in flight, but that doesn't seem
> useful.
> 
> Thus you have already lost all protection from the device, since your
> entire memory is accessible via the bypass mapping.
> 
> Which means what is the point of then having non-bypass mappings for
> other things ? Just to make things slower ?
> 
> I really don't see what this whole "bypass on a per-mapping basis" buys
> you.

In the Intel case, the mapping setup is entirely per-device (except for
crap devices and devices behind a PCIe-PCI bridge, etc.).

So you can happily have a passthrough mapping for *one* device, without
making that same mapping available to another device. You can make that
trade-off of speed vs. protection for each device in the system.

Currently we have the 'iommu=pt' option that makes us use passthrough
for *all* devices (except those whose dma_mask can't reach all of
physical memory). But I could live with something like Shamir is
proposing, which lets us do the bypass only for performance-sensitive
devices which actually *ask* for it (and of course we'd have a system-
wide mode which declines that request and does normal mappings anyway).

-- 
dwmw2