Re: PCI arbiter memory mapping

[Joan Lledó]

Hi,

I'm sorry I can't follow your discussion, I only know about the small 
part of the kernel I worked on.

El 16/8/21 a les 23:07, Sergey Bugaev ha escrit:
> I don't think I understand enough about the situation. It would help
> if you or Joan were to kindly give me some more context :)

Basically, libpciaccess gets a memory object at [1]. And later I need it 
in the arbiter at [2], to create a proxy over it.

To do that, the current code stores it in a structure I created called 
pci_user_data [3], and then it reads that structure from the arbiter [4].

> What's the issue you're trying to solve?

We are looking for another way to get the pager at [4] and get rid of 
that structure.

> As I understand it, there's the PCI arbiter, which is a translator
> that arbitrates access to PCI, which is a hardware bus that various
> devices can be connected to.

Yes, and the arbiter can play two roles: root arbiter, which uses x86 
module in libpciacces; and nested arbiter, which uses the hurd module in 
libpciaccess.

> The hardware devices connected via PCI are available (to the PCI arbiter)
> as Mach devices

Actually, the devices are available to the arbiter as libpciaccess devices

> it's possible to use device_map () and then vm_map () to access the
> device memory.

Yes, root arbiter uses device_map() on "mem" to get the memory object, 
nested arbiters use io_map() over the region files exposed by the root 
arbiter to get the memory object.

Both pass the memory object to vm_map() to map the range.

> Then there's libpciaccess whose Hurd backend uses the
> files exported by the PCI arbiter to get access to the PCI,

Only nested arbiters, as I said the root arbiter uses the x86 backend

> Naturally its user can request read-only or
> read-write mapping, but the PCI arbiter may decide to only return a
> read-only memory object (a proxy to the device pager), in which case
> libpciaccess should deallocate the port and return EPREM, or the PCI
> arbiter may return the real device pager.

Yes, but that's not really relevant for our problem, I was talking about 
a bug I found.


----------------
[1] 
https://gitlab.freedesktop.org/jlledom/libpciaccess/-/blob/hurd-device-map/src/x86_pci.c#L275
[2] 
http://git.savannah.gnu.org/cgit/hurd/hurd.git/tree/pci-arbiter/netfs_impl.c?h=jlledom-pci-memory-map#n613
[3] 
https://gitlab.freedesktop.org/jlledom/libpciaccess/-/blob/hurd-device-map/src/x86_pci.c#L287
[4] 
http://git.savannah.gnu.org/cgit/hurd/hurd.git/tree/pci-arbiter/netfs_impl.c?h=jlledom-pci-memory-map#n605