On Thu, Aug 18, 2022 at 10:59 PM Ben Hutchings <b...@decadent.org.uk> wrote: > > Control: retitle -1 [i386] Unconditional LFENCE instructions in > FILL_RETURN_BUFFER > Control: tag -1 confirmed upstream > Control: found -1 5.18.14-1 > > On Wed, 2022-08-17 at 11:42 +0200, Etienne Vogt wrote: > > I can confirm that this bug also occurs on Athlon XP systems (Generic VIA > > KT333 motherboard, CPU AMD Athlon(tm) XP 2600+) : kernel panic early on > > boot. > > > > I suspect someone thought it would be a good idea to compile the kernel > > for P4 only, as both PIII and Athlon XP processors lack the SSE2 > > instruction set. > > > > That was a good guess, though we don't change the configuration like > that in stable updates. > > The RETbleed mitigations, which are not needed on these CPUs or even > functional on 32-bit kernels, interact with the Spectre v2 mitigations, > which *are* used on these CPUs. And unfortunately the RETbleed > mitigations added some unconditional LFENCE instructions, which should > be conditional since they are part of SSE2. > > As a temporary workaround, disabling the Spectre v2 mitigation with the > kernel parameter "nospectre_v2" should allow this kernel version to run > on older CPUs without SSE2. We'll fix this properly in a later update.
That breakage affects Stable. Expecting people to go and use workarounds on what was meant to be a stable update isn't acceptable. I really hope that the fix will be pushed within the next 24 hours with high urgency. Martin-Éric
