Replace bus/slot/function with full PCI paths

[DemiMarie]

How to file a helpful issue

The problem you're addressing (if any)

PCI devices are currently assigned via bus/slot/function, which is not a stable identifier.

The solution you'd like

Use PCI paths instead, which are much more likely to be stable.

Specifically, replace the "bus" part with the bridges path starting from bus 0. For example:

# lspci -t
-[0000:00]-+-00.0
           +-00.2
           +-01.0
           +-02.0
           +-02.2-[01]----00.0
           +-02.4-[02]----00.0
           +-03.0
           +-03.1-[03-61]--
           +-04.0
           +-04.1-[62-c0]--
           +-08.0
           +-08.1-[c1]--+-00.0
           |            +-00.1
           |            +-00.2
           |            +-00.3
           |            +-00.4
           |            +-00.5
           |            \-00.6
...

The device 0000:c1:00.0 is behind bridge at 0000:00:08.1. In some cases there may be more bridges. There may be more devices behind a single bridge (in the example above it's just a single multi-function device).
You can get bus ranges handled by a given bridge by looking at "secondary" and "subordinate" attributes on lspci (and similarly in sysfs):

00:08.1 PCI bridge: Advanced Micro Devices, Inc. [AMD] Phoenix Internal GPP Bridge to Bus [C:A] (prog-if 00 [Normal decode])
	Subsystem: Device 0006:f111
	Flags: bus master, fast devsel, latency 0, IRQ 106
	Bus: primary=00, secondary=c1, subordinate=c1, sec-latency=0
...

I imagine the path for the 0000:c1:00.0 device would look something like this: 0000:00:08.1-00:00.0, where the -00 part means "the first bus behind this bridge". The important part is to not include "c1" in the address (but use the address relative to the bridge range), because if the bus will get renumbered (when somebody plugs some other devices for example), that number may change.

If there would be device at bus 5, (which is the third bus behind bridge at 00:03.1), for example at 0000:05:01.5, it would have a path like 0000:00:03.1-02:01.5

PCI bridges cannot cross the PCI segment boundaries, so the first "0000" will stay the same (doesn't need to be duplicated in the address).

The value to a user, and who that user might be

Users will be less likely to have problems with PCI device renumbering.

[jamke]

We had several discussions on this topic:

#7792
#7883

[marmarek]

Use PCI paths instead, which are much more likely to be stable.

Please be more specific what exactly you mean. sysfs path of the device? Can you give a specific example (that can be used as a device identifier, so meets qrexec service argument allowed character set)?

[DemiMarie]

Use PCI paths instead, which are much more likely to be stable.

Please be more specific what exactly you mean. sysfs path of the device?

Physical path to the device, as defined by the sequence of PCI switches that must be traversed. I think this corresponds to sysfs path.

Can you give a specific example (that can be used as a device identifier, so meets qrexec service argument allowed character set)?

One can turn the sysfs path into an identifier by escaping.

[DemiMarie]

Per Greg Kroah-Hartman on Mastodon it seems that slot IDs are known only to firmware, and the firmware might or might not provide one.

[marmarek]

Per Greg Kroah-Hartman on Mastodon it seems that slot IDs are known only to firmware, and the firmware might or might not provide one.

I don't think my proposal needs that info.

Physical path to the device, as defined by the sequence of PCI switches that must be traversed. I think this corresponds to sysfs path.

My proposal does more or less this. With one extra twist - that it's immune to bus renumbering. Specifically the identifier I propose will not change if somebody will plug something into another port. Sysfs path may change, as it includes absolute bus number at each step.

[DemiMarie]

What about device renumbering?

[marmarek]

The "device" part of the BDF is related to physical connection. I guess if you change the device (put into a different port an PCIe switch?) it may change, but that isn't a concern here.