On Fri, Oct 04, 2024 at 03:56:33PM +0200, Oleksij Rempel wrote: > On Thu, Oct 03, 2024 at 05:22:58PM +0200, Andrew Lunn wrote: > > > Indeed, but regulator API already provide such events, which will even be sent > > > when we enable or disable the PSE. Should we write a second event management. > > > Using regulator event API allows to report over current internal events to the > > > parents regulator the power supply of the PSE which could also do something to > > > avoid smoke. > > > > > > Or maybe we should add another wrapper which will send PSE ethtool netlink > > > notification alongside the regulator notifications supported by this patch. > > > > > > > Also, how do regulator events work in combination with network > > > > namespaces? If you move the interface into a different network > > > > namespace, do the regulator events get delivered to the root namespace > > > > or the namespace the interface is in? > > > > > > regulator events are sent in root namespace. > > > > I think we will need two event, the base regulator event, and a > > networking event. Since it is a regulator, sending a normal regulator > > event makes a lot of sense. But mapping that regulator event to a > > netns:ifnam is going to be hard. Anything wanting to take an action is > > probably going to want to use ethtool, and so needs to be in the > > correct netns, etc. But it does get messy if there is some sort of > > software driven prioritisation going on, some daemon needs to pick a > > victim to reduce power to, and the interfaces are spread over multiple > > network namespaces. > > > > What i don't know is if we can use an existing event, or we should add > > a new one. Often rtnetlink_event() is used: > > > > https://elixir.bootlin.com/linux/v6.12-rc1/source/net/core/rtnetlink.c#L6679 > > > > but without some PSE information in it, it would be hard to know why > > it was sent. So we probably either want a generic ethtool event, or a > > PSE event. > > Hm... assuming we have following scenario: > > .--------- PI 1 > / .--------- PI 2 > .========= PSE /----------( PI 3 ) NNS red > // \----------( PI 4 ) NNS blue > Main supply // `---------( PI 5 ) NNS blue > o================´--- System, CPU > > In this case we seems to have a new challenge: > > On one side, a system wide power manager should see and mange all ports. > On other side, withing a name space, we should be able to play in a > isolated sand box. There is a reason why it is isolated. So, we should > be able to sandbox power delivery and port prios too. Means, by creating > network names space, we will need a power names space. > > I can even imagine a use case: an admin limited access to a switch for > developer. A developer name space is created with PSE budget and max > prios available for this name space. This will prevent users from DoSing > system critical ports. > > At this point, creating a power name space will an overkill for this > patch set, so it should be enough to allow controlling prios over > ethtool per port and isolation support if needed. Oh, sorry, i'm too tired. Too many words are missing in my answer ... -- Pengutronix e.K. | | Steuerwalder Str. 21 | http://www.pengutronix.de/ | 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 | Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
