On 12/12/2019 1:12 PM, Michael Walle wrote:
Hi Alex,
Am 2019-12-11 22:01, schrieb Alexandru Marginean:
Hi Michael,
On 12/11/2019 6:03 PM, Michael Walle wrote:
Hi Alex,
Am 2019-12-11 16:37, schrieb Alexandru Marginean:
On 12/11/2019 2:16 PM, Michael Walle wrote:
Hi Vladimir,
Am 2019-12-11 13:46, schrieb Vladimir Oltean:
Hi Michael,
On Wed, 11 Dec 2019 at 00:48, Michael Walle michael@walle.cc wrote: > > Am 2019-12-10 15:55, schrieb Alex Marginean: > > Passes on the primary address used by u-boot to Linux. The > code does a > > DT > > fix-up for ENETC PFs and sets the primary MAC address in IERB. > The > > address > > in IERB is restored on ENETC PCI functions at FLR. > > > I don't get the reason why this is done in a proprietary way. > What is > the > difference between any other network interface whose hardware > address is > set up in the generic u-boot code. > > Also, shouldn't write_hwaddr callback be implemented instead of the > enetc_set_ierb_primary_mac()? >
At the moment, the Linux driver ignored the device tree and only reads the POR values of the SIPMAR registers. The reset value of those comes from the IERB space, which U-Boot is configuring. So it would be good if that behavior keeps working.
It would also be good if the Linux driver called of_get_mac_address, so it needs the device tree binding for that. That's why both fixups are performed, and why the generic function is not used.
yes, but u-boot already sets the mac-address/local-mac-address property in the device tree already in a generic way, see fdt_fixup_ethernet().
I think fdt_fixup_ethernet is not a good choice for us. The issue is that it ties Linux DT to device indexes in U-Boot. That's a problem if we plug an Eth PCI card in, we would need to change Linux DT, which is definitely not desirable. We actually do use PCI Eth cards with some of our boards and U-Boot does pick those up and indexes shift.
are you sure? afaik it works by reading the /alias/ethernetN phandles which gets ethNaddr assigned if you set the FDT_SEQ_MACADDR_FROM_ENV config option. I've just tried it, here is my linux dts diff
--- a/arch/arm64/boot/dts/freescale/fsl-ls1028a.dtsi +++ b/arch/arm64/boot/dts/freescale/fsl-ls1028a.dtsi @@ -17,6 +17,11 @@ #address-cells = <2>; #size-cells = <2>;
+ aliases { + ethernet0 = &enetc_port0; + ethernet1 = &enetc_port1; + };
cpus { #address-cells = <1>; #size-cells = <0>; @@ -761,10 +766,12 @@ enetc_port0: ethernet@0,0 { compatible = "fsl,enetc"; reg = <0x000000 0 0 0 0>; + local-mac-address = [ 00 00 00 00 00 00 ]; }; enetc_port1: ethernet@0,1 { compatible = "fsl,enetc"; reg = <0x000100 0 0 0 0>; + local-mac-address = [ 00 00 00 00 00 00 ]; }; enetc_mdio_pf3: mdio@0,3 { compatible = "fsl,enetc-mdio";
That way the mapping between ethNaddr and the network device can also be changed by the user by changing the linux device tree.
also, uboot should respect the /aliases/ethernetN, too.
I don't disagree with any of that, but that's not the issue I mentioned. I meant actual PCI cards being plugged in, I didn't mean having disabled ECAM functions.
In your example DT enetc port 0 is tied to /aliases/ethernet0 and to ethaddr, enetc port 1 is tied to /aliases/ethernet1 and to eth1addr.
A LS1028 board with a PCI Eth card plugged in shows this:
Net: e1000: 68:05:ca:66:bf:bd eth0: e1000#0 [PRIME], eth1: enetc-0, eth2: enetc-2, eth3: swp0, eth4: swp1, eth5: swp2, eth6: swp3
In this case eth0 is the e1000 card and it uses ethaddr, enetc port 0 is eth1 and uses eth1addr. The fix-up to /aliases/ethernet0 in your example makes enetc port 0 get the MAC address of the PCI card. If the PCI card is removed then enetc port 0 ends up being eth0 in U-Boot and and actually use ethaddr, not eth1addr.
Mh, I've just tried this and it seems that this is even worse. Because if you set the ethaddr, uboot will complain about mismatching ethernet addresses.
Between ethaddr from eeprom and e1000 ROM address? Yes, it will complain.
To make this work with a PCI card plugged in one would need to change the aliases in Linux DT, which is not a fun thing to do.
BTW what will be the source of the network addresses, the u-boot environment variables? (which might be set either by the user or some kind of board specific code).
Yes, the environment variables. As far as I know these come preset from the factory. The reference boards usually come with stickers too, listing the preset MAC addresses.
so here is already the first problem, see above. Just to be sure, I don't argue against having the environment as the source (actually we do import the mac addresses from our eeprom into the environment), just to make you aware that there is an underlying problem.
I know, we waste the 1st address from environment if there's a card plugged in, and having more cards makes it worse.
I've tried to use the aliases in u-boot, too. Unfortunately, the e1000 will get the first slot, regardless that /aliases/ethernet0 is a phandle to enetc-0. I guess that is the actual problem here. Because if that would work then everything else would fall into place. (except the write_hwaddr()).
I don't think aliases in U-Boot DT help really, these are all PCI devices, the card and enetc/ECAM. They are probed on the PCI bus, not from DT.
Also U-Boot and Linux DTs have to be in sync all the time, if we disable one port in U-Boot but not in Linux we would mix up MAC addresses.
I see. But that shouldn't happen with the code above. But are you sure that this
+ uclass_get(UCLASS_ETH, &uc); + uclass_foreach_dev(dev, uc) {
will work then? in my config (just enetc-0) there is only one eth device
# dm tree [snip] pci 2 [ + ] pci_generic_ecam |-- pcie@1f0000000 eth 0 [ + ] enetc_eth | |-- enetc-0 mdio 0 [ + ] enetc_mdio | |-- emdio-3 pci_generi 0 [ ] pci_generic_drv | |-- pci_4:0.4 dsa 0 [ ] felix-switch | |-- felix-switch pci_generi 1 [ ] pci_generic_drv | `-- pci_4:1f.0 [snip]
Yeah, I guess that's no better either way. U-Boot wouldn't be able to fix-up addresses for interfaces it didn't know about, yes. It works if Linux doesn't use some interfaces that U-Boot does, the MAC addresses per port would match between U-Boot and Linux. That case isn't too useful though, if they differ it's probably the other way around, and U-Boot has fewer interfaces.
There is no simple solution to satisfy all cases really.
mhh just to let you know some cases we have: (1) enetc-0, enetc-1 in u-boot enetc-0, enetc-1, enetc-2 (and switch) in linux (2) enetc-1 in u-boot all enetc's and switch in linux (3a) well the "standard" ones where linux == uboot (3b) "standard one" but just enetc-1, skipping enetc-0
OK, these look reasonable to me, but we need a new solution to make them work.
Using an index encoded in the DT to identify a given interface isn't OK, in my example there's a PCI card that gets probed first and then all IDs shift up.
correct, and i guess that is the actual problem here. If it would be appended to a pre-defined list (aka /aliases in u-boot) then there wouldn't be any shift.
These interfaces have to be probed from the PCI bus. Using /aliases to fix ip dev->seq may work for devices used by U-Boot, it's not so obvious how U-Boot would keep track of indexes of devices it doesn't know about though. I hope we can come up with something simpler than that.
Fixing up addresses for interfaces U-Boot doesn't know about is also a problem, as the fix-up code should stay clear of eth addresses used by live interfaces, on the assumption that those should continue using them. Assume the device has two interfaces, first one is disabled in U-Boot but enabled in Linux. Should it use ethaddr in Linux, or eth1addr? Ethaddr is associated with 2nd interface in U-Boot and it makes sense to maintain that association in Linux.
If you say ethNaddr always maps to /alias/ethernetN, you won't have a problem, will you? Other than ethXaddr might be unused.
Also consider the following: - you have consecutive mac addresses - u-boot uses enetc-1, enetc-0 is disabled, it would use ethaddr lets say (its ..:00) - linux uses enetc-0 and enetc-1, then if it is not shifted, enetc-0 will have (..:01) and enetc-1 stays with (..:00).
Although it doesn't impact the function, it still looks bad.
It doesn't look neat, but it's workable.
So as far as using generic fix-up code I'm all for it, but in this case we would need some platform specific rules to match Linux DT nodes to U-Boot eth addresses.
As for the write_hwaddr callback instead of enetc_set_primary_mac_addr, that is valid but I suppose it is outside the scope of this patch. That function is related to the runtime MAC address and not to the MAC passed to Linux.
according to the comment in eth-uclass.c this is not for (u-boot) runtime:
/* * Devices need to write the hwaddr even if not started so that Linux * will have access to the hwaddr that u-boot stored for the device. * This is accomplished by attempting to probe each device and calling * their write_hwaddr() operation. */
and the runtime mac address for u-boot is already set enetc_start().
-michael
This is fine, I'll move the IERB code to write_hwaddr.
This will also only be called if the device is not disabled in u-boot, from what I see. So it might be better to "fix" this.
I bet you are not the first/only one which needs this kind of "how do we propagate the hardware address to linux" where the devices might be unused in u-boot. There must be a better way to do this ;)
-michael
Yeah, I'm sure others bumped into this before. For sure fdt_fixup_ethernet as defined now can't cover all cases. It can only rely on the U-Boot interface index which is volatile with plug-in PCI cards. We could introduce some new index for integrated interfaces that's set by the driver, maybe improve on existing req_seq. In current code req_seq is not too useful, as dev->seq is allocated in order of discovery, so it doesn't help if 2nd found device requests seq 0, that's already in use. But that could probably be improved.
For now I think I'll stick to current code, proprietary as it is and with the limitation that U-Boot has to enable interfaces that Linux uses.
Please don't take that shortcut because it will be a pain to get it cleaned-up and retain backwards compatibility :( Also like explained above, it already falls short if you have ethaddr defined.
I am open to any better suggestions though.
see above and like you've described it. append any plugable interfaces at the end of the predefined list (with probably empty slots in used ethNaddr).
-michael
All solutions based on the interface index fall short, the index is really volatile if PCI cards are involved.
So here's two alternatives I can think of:
- have the index nailed in place for relevant interfaces, including the ones U-Boot doesn't even use. That could be done introducing a static range of indexes, which is not used for plug-in cards. Since this should be backward compatible we don't want to mess with IDs in use on current platforms, so we would need to move the static IDs up. Say there's a
#define ETH_STATIC_INDEX_BASE 0x10
enetc could do something like:
dev->req_seq = ETH_STATIC_INDEX_BASE + fno.
meaning PF0, 1, 2, 6 get eth indexes 0x10, 0x11, 0x12, 0x16.
The DT fix-up code could use this base when processing the Linux DT, so /aliases/ethernet0 gets eth10addr, /aliases/ethernet1 gets eth11addr and so on.
Cards like e1000 are left with 0 based index, so they don't interfere with this ID range if they are not too many, and MAC addresses between U-Boot and Linux are preserved even if U-Boot doesn't use all the interfaces. that, or
- move away for using the index to make this association. How about having an allocation scheme that associates the device, not its probing index with the MAC address. It could be based on the name, like 'enetc-0-addr=..:00' or device topology, bus/function for pci, like 'eth-pci:4:00.0-addr=..:00'. These kind of schemes would not depend on what is enabled in U-Boot either, but they would require some more specific support for matching enetc-0 or pci:4:00.0 to a Linux DT node.
Thoughts, any better ideas?
Thanks! Alex