[sled-agent] Monitor for Tofino driver as factor in 'are we scrimlet' decision #1918
Conversation
sled-agent/src/sled_agent.rs
Outdated
|
|
||
| // Scan the existing system for noteworthy events | ||
| // that may have happened before we started monitoring | ||
| if self.inner.hardware.is_scrimlet() { |
There was a problem hiding this comment.
Finding an attached tofino driver tells us that we're running on a scrimlet, but it doesn't mean we have all the resources needed to launch a switch zone. Specifically, we need the tfpkt0 link to exist. Without that, tfportd can't create tfports, so mgs and maghemite can't run. In practice, I would expect that link to exist long before sled-agent runs, but it's at least a theoretical concern.
There was a problem hiding this comment.
I fixed this in #1933 , when actually launching the switch zone
andrewjstone
left a comment
andrewjstone
left a comment
There was a problem hiding this comment.
Sean this looks good! Mostly minor comments from me. Based on @davepacheco's comments and my rethinking about the nexus queue in #1917, I'm not sure we want to stick with that strategy. However, I'm fine merging this as is so we can start using it and revisit that strategy in follow ups with whatever we decide upon.
| /// as illumos systems. | ||
| pub struct Hardware { | ||
| log: Logger, | ||
| inner: Mutex<HardwareInner>, |
There was a problem hiding this comment.
Is this Mutex actually needed? It looks like scrimlet config is only set during initialization.
There was a problem hiding this comment.
Admittedly, this isn't really needed. I was trying to emulate the OPTE usage here: https://github.com/oxidecomputer/omicron/tree/main/sled-agent/src/opte
... But in reality, the Sled Agent only support the "sim" version on non-illumos platforms. If we try running the "real" sled agent outside illumos, tons of stuff will break (access to zones, dladm, ipadm, etc).
I've decided to keep most of this module as stubs, so that editor support (like rust analyzer) will still work on platforms like Linux, but it'll hopefully be more obvious that this code should not be callable when we start approaching the "real illumos" interfaces. (done in 4863667)
andrewjstone
left a comment
andrewjstone
left a comment
There was a problem hiding this comment.
Looks great! Thanks for the cleanup and comments.
hardwaremodule to Sled Agent for minimally monitoringdevinfooutput. This will certainly evolve, but this PR includes a "bare minimum" real example of tracking the tofino driver.rss-config.tomlto decide if we're running on a scrimlet. Instead...force-scrimletoption to make the sled agent assume that it is a scrimletFixes https://github.com/oxidecomputer/minimum-upgradable-product/issues/19
Part of #1917
Part of #823
Pre-requisite for https://github.com/oxidecomputer/minimum-upgradable-product/issues/16
Pre-requisite for https://github.com/oxidecomputer/minimum-upgradable-product/issues/18