| kata-irc-bot | <dgibson> The current approach isn't great, but fixing it isn't trivial | 02:20 |
|---|---|---|
| kata-irc-bot | <dgibson> the reason for the bridge is that it gives a whole bunch of slots for a fairly low cost | 02:20 |
| kata-irc-bot | <dgibson> to give the equivalent number of available slots we'd need a lot of pcie ports, which costs resources | 02:21 |
| kata-irc-bot | <dgibson> or kind of chain pcie ports and switches, which is much fiddlier to implement | 02:21 |
| kata-irc-bot | <dgibson> and has possible complications with the top-level port having IO windows constrained by things way down the chain | 02:21 |
| kata-irc-bot | <dgibson> in addition, pcie ports currently use pcie native hotplug which, unfortunately, has some problems on the guest side | 02:22 |
| kata-irc-bot | <dgibson> it can race with the guest's initial probe leading to devices not being plugged properly | 02:22 |
| kata-irc-bot | <dgibson> that should be fixable, but again, it's not straightforward | 02:22 |
| kata-irc-bot | <dgibson> we could switch to ACPI hotplug even for pcie ports | 02:23 |
| kata-irc-bot | <dgibson> but that's probably less well tested on the guest side, and won't work for ARM or other platforms | 02:23 |
| kata-irc-bot | <dgibson> or we could try to fix the races in the kernel | 02:23 |
| kata-irc-bot | <dgibson> or we can try to avoid them from Kata, but it's not entirely clear how to do that | 02:23 |
| kata-irc-bot | <treva930> I was at that session :@ | 14:30 |
| kata-irc-bot | <fidencio> Folks, as I'm going for vacation from Friday, I'd like to ask if someone would have cycles to review: https://github.com/kata-containers/kata-containers/pull/4540 It's a not too complex, not too intrusive PR, and we agreed on having this one in before the -rc0 release. | 21:10 |
Generated by irclog2html.py 2.17.3 by Marius Gedminas - find it at https://mg.pov.lt/irclog2html/!