On Tue Nov 28, 2023 at 10:59 AM AEST, Michael Ellerman wrote: > Christophe Leroy <christophe.le...@csgroup.eu> writes: > > Le 27/11/2023 à 19:39, Timothy Pearson a écrit : > >> Just wanted to check back and see if this patch was going to be > >> queued up soon? We're still having to work around / advertise the > >> data destruction issues the underlying bug is causing on e.g. Debian > >> Stable. > > > > Has any agreement been reach on the final solution ? Seeing the many > > discussion on patch v2 I had the feeling that it was not the final solution. > > The actual patch is fine I think. > > The discussion was about improving the explanation of exactly what's > happening in the change log, and whether there is a larger bug causing > FP corruption unrelated to io-uring.
One thing I said is maybe we should remove the "register save" concept entirely and always giveup. But that would not be suitable for a minimal fix. I didn't mean it as an alternate fix. Even if we did always giveup in future, this patch still gives a nicer API. There would have to be a noticable performance advantage to not restoring fr0/vs0 after saving before we'd think about changing it back to clobber, since that encourages foot shooting. Thanks, Nick