Re: [PATCH] arm64: dts: marvell: armada-37xx: Set linux,pci-domain to zero

From: Pali Rohár
Date: Sat Apr 17 2021 - 10:50:07 EST


On Thursday 15 April 2021 10:13:17 Rob Herring wrote:
> On Thu, Apr 15, 2021 at 3:45 AM Marek Behun <marek.behun@xxxxxx> wrote:
> >
> > On Thu, 15 Apr 2021 10:36:40 +0200
> > Pali Rohár <pali@xxxxxxxxxx> wrote:
> >
> > > On Tuesday 13 April 2021 13:17:29 Rob Herring wrote:
> > > > On Mon, Apr 12, 2021 at 7:41 AM Pali Rohár <pali@xxxxxxxxxx> wrote:
> > > > >
> > > > > Since commit 526a76991b7b ("PCI: aardvark: Implement driver 'remove'
> > > > > function and allow to build it as module") PCIe controller driver for
> > > > > Armada 37xx can be dynamically loaded and unloaded at runtime. Also driver
> > > > > allows dynamic binding and unbinding of PCIe controller device.
> > > > >
> > > > > Kernel PCI subsystem assigns by default dynamically allocated PCI domain
> > > > > number (starting from zero) for this PCIe controller every time when device
> > > > > is bound. So PCI domain changes after every unbind / bind operation.
> > > >
> > > > PCI host bridges as a module are relatively new, so seems likely a bug to me.
> > >
> > > Why a bug? It is there since 5.10 and it is working.
>
> I mean historically, the PCI subsystem didn't even support host
> bridges as a module. They weren't even proper drivers and it was all
> arch specific code. Most of the host bridge drivers are still built-in
> only. This seems like a small detail that was easily overlooked.
> unbind is not a well tested path.

Ok! Just to note that during my testing I have not spotted any issue.

> > > > > Alternative way for assigning PCI domain number is to use static allocated
> > > > > numbers defined in Device Tree. This option has requirement that every PCI
> > > > > controller in system must have defined PCI bus number in Device Tree.
> > > >
> > > > That seems entirely pointless from a DT point of view with a single PCI bridge.
> > >
> > > If domain id is not specified in DT then kernel uses counter and assigns
> > > counter++. So it is not pointless if we want to have stable domain id.
> >
> > What Rob is trying to say is that
> > - the bug is that kernel assigns counter++
> > - device-tree should not be used to fix problems with how kernel does
> > things
> > - if a device has only one PCIe controller, it is pointless to define
> > it's pci-domain. If there were multiple controllers, then it would
> > make sense, but there is only one
>
> Yes. I think what we want here is a domain bitmap rather than a
> counter and we assign the lowest free bit. That could also allow for
> handling a mixture of fixed domain numbers and dynamically assigned
> ones.

Currently this code is implemented in pci_bus_find_domain_nr() function.
IIRC domain number is 16bit integer, so plain bitmap would consume 8 kB
of memory. I'm not sure if it is fine or some other tree-based structure
for allocated domain numbers is needed.

> You could create scenarios where the numbers change on you, but it
> wouldn't be any different than say plugging in USB serial adapters.
> You get the same ttyUSBx device when you re-attach unless there's been
> other ttyUSBx devices attached/detached.

This should be fine for most scenarios. Dynamically attaching /
detaching PCI domain is not such common action...

Will you implement this new feature?