Re: [RFC PATCH net-next 1/8] ethtool: Add ability to control transceiver modules' low power mode
From: Ido Schimmel <hidden>
Date: 2021-08-10 20:46:36
On Tue, Aug 10, 2021 at 06:59:54AM -0700, Jakub Kicinski wrote:
On Tue, 10 Aug 2021 15:52:20 +0200 Andrew Lunn wrote:quoted
quoted
The transition from low power to high power can take a few seconds with QSFP/QSFP-DD and it's likely to only get longer with future / more complex modules. Therefore, to reduce link-up time, the firmware automatically transitions modules to high power mode. There is obviously a trade-off here between power consumption and link-up time. My understanding is that Mellanox is not the only vendor favoring shorter link-up times as users have the ability to control the low power mode of the modules in other implementations. Regarding "why do we need user space involved?", by default, it does not need to be involved (the system works without this API), but if it wants to reduce the power consumption by setting unused modules to low power mode, then it will need to use this API.O.K. Thanks for the better explanation. Some of this should go into the commit message. I suggest it gets a different name and semantics, to avoid confusion. I think we should consider this the default power mode for when the link is administratively down, rather than direct control over the modules power mode. The driver should transition the module to this setting on link down, be it high power or low power. That saves a lot of complexity, since i assume you currently need a udev script or something which sets it to low power mode on link down, where as you can avoid this be configuring the default and let the driver do it.Good point. And actually NICs have similar knobs, exposed via ethtool priv flags today. Intel NICs for example. Maybe we should create a "really power the port down policy" API?
See below about Intel. I'm not sure it's the same thing... I'm against adding a vague "really power the port down policy" API. The API proposed in the patch is well-defined, its implementation is documented in standards, its implications are clear and we offer APIs that give user space full observability into its operation. A vague API means that it is going to be abused and user space will get different results over different implementations. After reading the *commit messages* about the private flags, I'm not sure what the flags really do, what is their true motivation, implications or how do I get observability into their operation. I'm not too hopeful about the user documentation. Also, like I mentioned in the cover letter, given the complexity of these modules and as they become more common, it is likely that we will need to extend the API to control more parameters and expose more diagnostic information. I would really like to keep it clean and contained in 'ETHTOOL_MSG_MODULE_*' messages and not spread it over different APIs.
Jake do you know what the use cases for Intel are? Are they SFP, MAC, or NC-SI related?
I went through all the Intel drivers that implement these operations and I believe you are talking about these commits: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=c3880bd159d431d06b687b0b5ab22e24e6ef0070 https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=d5ec9e2ce41ac198de2ee18e0e529b7ebbc67408 https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=ab4ab73fc1ec6dec548fa36c5e383ef5faa7b4c1 There isn't too much information about the motivation, but maybe it has something to do with multi-host controllers where you want to prevent one host from taking the physical link down for all the other hosts sharing it? I remember such issues with mlx5.
quoted
I also wonder if a hierarchy is needed? You can set the default for the switch, and then override is per module? I _guess_ most users will decide at a switch level they want to save power and pay the penalty over longer link up times. But then we have the question, is it an ethtool option, or a devlink parameter?