On Wed, Sep 25, 2024 at 08:48:31PM +0200, Christophe Leroy wrote: > > > Le 25/09/2024 à 20:38, Jason A. Donenfeld a écrit : > > On Wed, Sep 25, 2024 at 07:50:22PM +0200, Jason A. Donenfeld wrote: > >> For gettimeofday, -ffixed-r30 was passed to work around a bug in Go > >> code, where the vDSO trampoline forgot to save and restore this register > >> across function calls. But Go requires a different trampoline for every > >> call, and there's no reason that new Go code needs to be broken and add > >> more bugs. So remove -ffixed-r30 for getrandom. > > > > Strangely, I am _unable to_ make the Go code not crash with this patch > > applied. I'm not quite sure what I'm doing wrong yet, or if this points > > to another issue. > > Do you mean that without this patch the Go code works and with this > patch it crashes ? That's strange taken into account that the chacha > function plays up with r30 regardless of this patch.
Yea, that's what I meant. It turned out to be a Go runtime bug, which won't affect other vDSO functions but does affect getrandom(). I fixed the issue and sent a patch up to Google. So this is not the kernel's issue. Therefore, this patch can move forward. Jason
