On Tue, Dec 11, 2018 at 06:48:16PM +0000, Dave Martin wrote: > Currently FPEXC32_EL2 is handled specially when context-switching. > > However, FPEXC has no architectural effect when running in AArch64. > The only case where an arm64 host may execute in AArch32 is when > running compat user code at EL0: the architecture explicitly > documents FPEXC as having no effect in this case either when the > kernel (EL2 in the VHE case; otherwise EL1) is AArch64. > > So, we can just handle FPEXC32_EL2 (which is the AArch64 alias for > this register) as a regular AArch32-only system register and switch > it without any special handling or synchronisation. > > This allows some gnarly special-case code to be eliminated from > hyp. The extra isb() in __activate_traps_fpsimd32() is dropped > too: for the reasons explained above arm64 never required this > anyway. > > Note that the comment in the deleted __activate_traps_fpsimd32() > function about the need to inhibit traps to EL1 is only meaningful > if the guest's value for FPEXC is not restored (thus potentially > causing an unexpected trap to EL1 in the guest unless we're careful > to set the EN bit). This arises from the treatment of FPEXC32_EL2 > as if it were part of the the FPSIMD state to be switched lazily, > whereas it is really more natural to consider it as part of the EL1 > system register state to be switched eagerly at the appropriate > times. > > In fact, we explicitly don't care about seeing any trap until the > guest accesses the FPSIMD registers for real, provided that the > guest sees the behaviour it expects. So ensuring that the guest's > view of FPEXC is alawys up to date (as implemented by this patch) > should be sufficient. Since we always had to write FPEXC32_EL2 > anyway (in order to work around the possibility of not restoring it > properly for the guest) this should not add significant cost. In > any case, this patch moves any cost off the hyp switch critical > path, to vcpu_load/put(). > > Cc: Marc Zyngier <[email protected]> > Cc: Christoffer Dall <[email protected]> > Cc: Alexander Graf <[email protected]> > Signed-off-by: Dave Martin <[email protected]> > > --- > > **Build-tested only**
Dang, it turns out that writing FPEXC32_EL2 from EL2 triggers a trap to EL2 if FPSIMD traps to EL2 are enabled via CPTR_EL2. This seems illogical, but I've seen it trigger a Hyp panic on Juno (r0) (non-VHE, but looking at the Arm ARM it appears that something equivalent happens for VHE too). So this needs a rethink :| I think we can still defer the save until vcpu_put(), and save a sysreg write by not restoring it lazily (since we have to write it once anyway to set EN bit if we want to trap FPSIMD to EL2). Anyway, definitely not a priority for v4.21. [...] Cheers ---Dave _______________________________________________ kvmarm mailing list [email protected] https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
