Re: PCI Arbiter

[Samuel Thibault]

Hello,

Joan Lledó, on dim. 03 déc. 2017 11:21:53 +0100, wrote:
> This is a new version of the PCI arbiter.

Nice work :D

> I tried to run netdde as non-root but gave up because it seemed non-trivial 
> for
> me and the time for my thesis is running out. Next are some issues I found:
> 
> - Some network drivers poke IO ports by themselves (See #hurd, 2017-11-23 
> [1]).

Yes. Thinking about it, Mach actually provides a way to give a task the
right for a given range of I/O ports (i386_io_perm_cerate), so one could
use that.

> - check_kernel() in check_kernel.c[2] is calling get_privileged_ports().

We could make this just return if it's not privileged. Longterm-wise we
want to just drop kernel network drivers anyway.

> Besides, the pci server lacks some features required to run netdde as 
> non-root:
> 
> - Provide some way for the client to map device's regions and rom into its
>   space. read() + mmap() could be enough for read-only spaces, but probably
>   making the server act as a pager is the only good solution.
>   (See #hurd, 2017-10-27 [3]).

Possibly yes. Virtual memory hackers, any opinion on this?

> - Provide libpciaccess and pciutils with a way to poke IO ports as non-root.
>   Adding new routines to the PCI interface is an option, or even creating a
>   new interface, since poking IO ports is not necessarily related to PCI.

Well, as mentioned above the interface already exists actually :) But
there doesn't exist any arbiter for it in general. I don't think we want
to spend time on doing it for legacy ports, and we should just work on
support in the PCI arbiter: an RPC operation which just returns a port
created with i386_io_perm_create, so the caller can then enable/disable
I/O access to the ranges of ports at will.

Samuel