Re: [pve-devel] [PATCH RFC container] Add device passthrough

[Dominik Csapak <d.csapak@proxmox.com>]

On 10/20/23 10:29, Thomas Lamprecht wrote:
> Am 20/10/2023 um 09:51 schrieb Dominik Csapak:
>> On 10/20/23 09:08, Wolfgang Bumiller wrote:
>>> Also, Dominik recently added resource mappings for qemu for USB & PCI.
>>> PCI might be tricky, but for USB we may be able to use these mappings as
>>> well.
>>> That said, "raw" `/dev` node pass-through still makes sense as a
>>> separate thing for containers anyway since raw `lxc....` entries in the
>>> container config can currently be very inconvenient to deal with
>>> particularly with unprivileged containers (read on below for why...)
>>
>> just to chime in here, i don't think it'll be easily possible to reuse
>> the pci/usb maps as is since we'd have to map from pciid /usb-vendor/device
>> (or path) to a device node? i don't think thats generally possible, since
>> the driver does not always make that info easily available
>> (e.g. multi gpu setup and /dev/dri/cardX, or usb-to-serial adapters
>> and /dev/ttySX ?) i guess it could work, but we probably would have
>> to implement that for every driver out there?
>>
> 
> USB should be workable via resolving to /dev/bus/usb/*, PCI could be,
> theoretically, but isn't now and probably won't be anytime soon – i.e.,
> as Wolfgang mentioned off list, there's a reason that there's no
> /dev/bus/pci/
> 
>> what i would like to see however is to integrate a new type of mapping
>> for container devices specifically so that the ux is the same
>> (create mappings for whole cluster, assigning privileges, etc)
> 
> I'd try hard to re-use the USB mappings, those seem to be one of the
> most common pass-through setups for containers (e.g., for those
> home automation zigbee/matter/... adapters, or in some countries DVB-T
> ones, be it for TV or ADS-B plane tracking).

i guess, but sadly the /dev/bus/usb endpoint is mostly not what you want
to pass-through but the driver specific /dev/ttySX /dev/dvb/X and so on
(there are situations where the /dev/bus/usb path is the wanted one,
but there are many where it isn't)

and while we can map from those to the usb device/vendor/path the reverse
mapping is not so easy (at least when i tried i did not found a generic way
via udev or similar

i would welcome it though if there is a way to do that of course

> 
> If we can make USB work then we'd have the basic concept of attaching
> a mapping done, adding a new type of (block/char) device mapping could
> then be an independent task for later to keep scope a bit smaller.
> 
> Fixing Wolfgang's comments for workable pass-through for unprivileged
> containers is probably the most important change needed for now.
> 
> I'd then even be open to apply this with (root@pam only!) absolute
> path to /dev support only, but IMO resolving the mapping itself should
> not be too hard, so if using /dev/bus/usb/ works having that in there
> from the start would be definitively nice.