Re: How to correctly define memory range of PCIe config space

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tuesday 09 August 2022 11:06:18 Rob Herring wrote:
> On Tue, Aug 9, 2022 at 10:29 AM Pali Rohár <pali@xxxxxxxxxx> wrote:
> >
> > Hello!
> >
> > On Tuesday 09 August 2022 09:59:49 Rob Herring wrote:
> > > On Sat, Aug 6, 2022 at 5:17 AM Pali Rohár <pali@xxxxxxxxxx> wrote:
> > > >
> > > > On Saturday 06 August 2022 16:36:13 Manivannan Sadhasivam wrote:
> > > > > Hi Pali,
> > > > >
> > > > > On Mon, Jul 11, 2022 at 12:51:08AM +0200, Pali Rohár wrote:
> > > > > > Hello!
> > > > > >
> > > > > > Together with Mauri we are working on extending pci-mvebu.c driver to
> > > > > > support Orion PCIe controllers as these controllers are same as mvebu
> > > > > > controller.
> > > > > >
> > > > > > There is just one big difference: Config space access on Orion is
> > > > > > different. mvebu uses classic Intel CFC/CF8 registers for indirect
> > > > > > config space access but Orion has direct memory mapped config space.
> > > > > > So Orion DTS files need to have this memory range for config space and
> > > > > > pci-mvebu.c driver have to read this range from DTS and properly map it.
> > > > > >
> > > > > > So my question is: How to properly define config space range in device
> > > > > > tree file? In which device tree property and in which format? Please
> > > > > > note that this memory range of config space is PCIe root port specific
> > > > > > and it requires its own MBUS_ID() like memory range of PCIe MEM and PCIe
> > > > > > IO mapping. Please look e.g. at armada-385.dtsi how are MBUS_ID() used:
> > > > > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/arm/boot/dts/armada-385.dtsi
> > > > > >
> > > > >
> > > > > On most of the platforms, the standard "reg" property is used to specify the
> > > > > config space together with other device specific memory regions. For instance,
> > > > > on the Qcom platforms based on Designware IP, we have below regions:
> > > > >
> > > > >       reg = <0xfc520000 0x2000>,
> > > > >             <0xff000000 0x1000>,
> > > > >             <0xff001000 0x1000>,
> > > > >             <0xff002000 0x2000>;
> > > > >       reg-names = "parf", "dbi", "elbi", "config";
> > > > >
> > > > > Where "parf" and "elbi" are Qcom controller specific regions, while "dbi" and
> > > > > "config" (config space) are common to all Designware IPs.
> > > > >
> > > > > These properties are documented in: Documentation/devicetree/bindings/pci/qcom,pcie.yaml
> > > > >
> > > > > Hope this helps!
> > > >
> > > > Hello! I have already looked at this. But as I pointed in above
> > > > armada-385.dtsi file, mvebu is quite complicated. First it does not use
> > > > explicit address ranges, but rather macros MBUS_ID() which assign
> > > > addresses at kernel runtime by mbus driver. Second issue is that config
> > > > space range (like any other resources) are pcie root port specific. So
> > > > it cannot be in pcie controller node and in pcie devices is "reg"
> > > > property reserved for pci bdf address.
> > > >
> > > > In last few days, I spent some time on this issue and after reading lot
> > > > of pcie dts files, including bindings and other documents (including
> > > > open firmware pci2_1.pdf) and I'm proposing following definition:
> > > >
> > > > soc {
> > > >   pcie-mem-aperture = <0xe0000000 0x08000000>; /* 128 MiB memory space */
> > > >   pcie-cfg-aperture = <0xf0000000 0x01000000>; /*  16 MiB config space */
> > > >   pcie-io-aperture  = <0xf2000000 0x00100000>; /*   1 MiB I/O space */
> > > >
> > > >   pcie {
> > > >     ranges = <0x82000000 0 0x40000     MBUS_ID(0xf0, 0x01) 0x40000  0x0 0x2000>,    /* Port 0.0 Internal registers */
> > > >              <0x82000000 0 0xf0000000  MBUS_ID(0x04, 0x79) 0        0x0 0x1000000>, /* Port 0.0 Config space */
> > >
> > > IMO, this should be 0 for first cell as this is config space. What is
> > > 0xf0000000 as that's supposed to be an address in PCI address space.
> >
> > Which value should be 0? I did not catch it here.
> 
> s/0x82000000/0/
> 
> But really that's 0 for the hi byte and BDF for the lower bytes.

But then, this line would describe "PCI device", not the address range
(if we should follow that pci2_1.pdf).

Also note that kernel accept only 0xX2000000 and 0xX1000000 values (MEM
and IO).

> > 0xf0000000 is just start offset for the child address range. I chose it
> > "randomly" to not conflict with 0x40000 offset for child address range
> > which describe internal PCIe controller registers.
> >
> > > >              <0x82000000 1 0x0         MBUS_ID(0x04, 0x59) 0        0x1 0x0>,       /* Port 0.0 Mem */
> > > >              <0x81000000 1 0x0         MBUS_ID(0x04, 0x51) 0        0x1 0x0>,       /* Port 0.0 I/O */
> > >
> > > I/O space at 4GB? It's only 32-bits. I guess this is already there
> > > from the mvebu binding, but it seems kind of broken...
> >
> > You have not looked how it works. mvebu is 32-bit CPU and this describe
> > whole address range which could possibly used for IO or MEM for any
> > (unspecified) PCIe controller.
> 
> I did look. It's just an intermediate virtual/random/made-up address
> space. It kind of follows PCI bus addressing, but then not really.

Exactly!

> > Every PCIe controller has its own PCIe MEM and PCIe IO address range
> > which are not shared with other PCIe controllers. (non-mvebu platforms
> > use for this different PCI domains; mvebu not)
> 
> Except there's this shared 'PCI like' bus each port is on.

I mean that in HW.

> > These "shared" ranges are then dynamically split into PCIe controllers
> > like their drivers ask. So at the end all 10 PCIe controllers which are
> > on AXP ask for maximum (64 kB) of PCIe IO and allocator then choose 10
> > 64kB non-interleaving ranges from this "big" master range. Same for PCIe
> > MEM.
> >
> > So basically those two lines (one for MEM and one for IO) just says that
> > "unspecified" amount of PCIe MEM is mapped to MBUS_ID(0x04, 0x59) and
> > another "unspecified" amount of PCIe IO is mapped to MBUS_ID(0x04, 0x51)
> > mbus and pci-mvebu.c drivers than takes care for proper allocation of
> > required memory from dynamic pool defined in pcie-mem-aperture and
> > pcie-io-aperture properties.
> 
> That sounds shared to me...

In device tree it looks like shared, but in HW it is not shared.

> >
> > Note that this is not something which I invented, this is used for a
> > long time by mvebu drivers and device tree files.
> >
> > > >
> > > >     pcie@1,0 {
> > > >       reg = <0x0800 0 0 0 0>; /* BDF 0:1.0 */
> > > >       assigned-addresses =     <0x82000800 0 0x40000     0x0 0x2000>,     /* Port 0.0 Internal registers */
> > > >                                <0x82000800 0 0xf0000000  0x0 0x1000000>;  /* Port 0.0 Config space */
> > >
> > > This says it is memory space, not config space.
> >
> > "config space" for Orion (in from driver point of view) is just ordinary
> > "memory space".
> 
> "memory space" means PCI memory space which it is not. Almost every
> Arm platform has config space in ordinary host/cpu address space.

Yes. This is not PCI (memory space). And it applies also for internal
registers range, which is prevent in all existing mvebu dts files.

Both assigned-addresses in above example (and also in all existing mvebu
dts files) describes host/cpu address space. Not PCI address space.

> > "memory space" is accessed by load/store instructions. I/O space is
> > accessed by inb/outb (IO instructions or what architecture has) and
> > "config space" is accessed by kernel drivers by its own drivers.
> >
> > So I think it is correct to declare this range of address space as
> > "memory".
> 
> It is not if we are talking PCI addresses. If this is Marvell pseudo
> virtual PCI addresses, then you can declare it whatever you want.

Yes, lets say it is "marvell pseudo pci address" which is not pci
address at all. But is of type "memory", not "io".

> > > But the PCI binding
> > > says config space is not allowed in assigned-addresses.
> >
> > Exactly. And this is because it does not make sense to assign "config
> > space" into "assigned-addresses" as described in:
> > https://www.devicetree.org/open-firmware/bindings/pci/pci2_1.pdf
> >
> > My understanding of above document is that "reg" property in following
> > code describes "config space".
> >
> >   pcie@1,0 {
> >     reg = <0x0800 0 0 0 0>; /* BDF 0:1.0 */
> >   }
> >
> > And reason is that because above document use "config space" terminology
> > for describing specific PCI device on the bus.
> >
> > Address part after the at-char (@1,0) in device tree should match "reg"
> > property and by all above definitions from pci2_1.pdf this pass because:
> >
> > 1) that document describes @1,0 format that is "config space" for PCI
> >    device with BDF X:1.0 where X is the parent bus
> >
> > 2) 0x0800 is really describes BDF X:1.0
> >
> > And then the "config space is not allowed in assigned-addresses" makes
> > sense as there is absolutely no need to put "pointer" to PCI device
> > (with BDF address) into "assigned-addresses" property. In this property
> > should be some memory range and not reference to some PCI device.
> >
> > > I think the parent ranges needs a config space entry with the BDF for
> > > each root port and then this entry should be dropped. It really looks
> > > to me like the mvebu binding created these fake PCI addresses to map
> > > root ports back to MBUS addresses when BDF could have been used
> > > instead.
> > >
> > > Rob
> >
> > The reason is that in mvebu there are X independent single-root-port
> > PCIe controllers (PCIe host bridges; PCI domains; ...) and device tree
> > binding was defined that all root ports (with their host bridges) would
> > be put into one "virtual" DT node which would act as one PCI domain.
> 
> If they are independent, why is there 1 mem and i/o aperture?
> 
> I don't know how a virtual DT node for purposes of creating a single
> PCI domain ever passed review. Must have been explained differently at
> the time.

Well, I do not know. I took that pci mvebu code as it was. I spend lot
of time understanding how is that marvell pcie hw working and how it is
implemented in linux kernel drivers and how it is defined in device
tree.

And I already saw that even marvell people did not understand how their
pcie hw is working; so it possible that nobody had all architecture and
hw information about this design at the time of device tree review.

Anyway, this is something which already exists, it is used and cannot be
changed for backward compatibility.

> > And this leads to the fact that all controller specific settings must be
> > defined in pci root ports.
> 
> Not necessarily. There's no reason settings can't be in the parent node.
> 
> Rob

Well, you can put it also into the parent node, but then you need to
somehow say which internal registers range describe which pcie root port
(=pcie controller). For me the logical option is to describe it is in
root port (=pcie controller) node as it the node, which needs it in hw.

But also this is how it is already used.



[Index of Archives]     [DMA Engine]     [Linux Coverity]     [Linux USB]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [Greybus]

  Powered by Linux