These patches fix some issues with the way KVM manages FPSIMD/SVE/SME state. The series supersedes my earlier attempt at fixing the host SVE state corruption issue:
https://lore.kernel.org/linux-arm-kernel/20250121100026.3974971-1-mark.rutla...
Patch 1 addresses the host SVE state corruption issue by always saving and unbinding the host state when loading a vCPU, as discussed on the earlier patch:
https://lore.kernel.org/linux-arm-kernel/Z4--YuG5SWrP_pW7@J2N7QTR9R3/ https://lore.kernel.org/linux-arm-kernel/86plkful48.wl-maz@kernel.org/
Patches 2 to 4 remove code made redundant by patch 1. These probably warrant backporting along with patch 1 as there is some historical brokenness in the code they remove.
Patches 5 to 7 are preparatory refactoring for patch 8, and are not intended to have any functional impact.
Patch 8 addresses some mismanagement of ZCR_EL{1,2} which can result in the host VMM unexpectedly receiving a SIGKILL. To fix this, we eagerly switch ZCR_EL{1,2} at guest<->host transitions, as discussed on another series:
https://lore.kernel.org/linux-arm-kernel/Z4pAMaEYvdLpmbg2@J2N7QTR9R3/ https://lore.kernel.org/linux-arm-kernel/86o6zzukwr.wl-maz@kernel.org/ https://lore.kernel.org/linux-arm-kernel/Z5Dc-WMu2azhTuMn@J2N7QTR9R3/
The end result is that KVM loses ~100 lines of code, and becomes a bit simpler to reason about.
I've pushed these patches to the arm64-kvm-fpsimd-fixes-20250210 tag on my kernel.org repo:
https://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git/ git://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git
The (unstable) arm64/kvm/fpsimd-fixes branch in that repo contains the fixes plus additional debug patches I've used for testing. I've given this some basic testing on a virtual platform, booting a host and a guest with and without constraining the guest's max SVE VL, with:
* kvm_arm.mode=vhe * kvm_arm.mode=nvhe * kvm_arm.mode=protected (IIUC this will default to hVHE)
Since v1 [1]: * Address some additional compiler warnings in patch 7 * Use ZCR_EL1 alias in VHE code * Fold in Tested-by and Reviewed-by tags * Fix typos
Since v2 [2]: * Ensure context synchronization in patch 8 * Fold in Tested-by and Reviewed-by tags * Fix typos
[1] https://lore.kernel.org/linux-arm-kernel/20250204152100.705610-1-mark.rutlan... [2] https://lore.kernel.org/linux-arm-kernel/20250206141102.954688-1-mark.rutlan...
Mark.
Mark Rutland (8): KVM: arm64: Unconditionally save+flush host FPSIMD/SVE/SME state KVM: arm64: Remove host FPSIMD saving for non-protected KVM KVM: arm64: Remove VHE host restore of CPACR_EL1.ZEN KVM: arm64: Remove VHE host restore of CPACR_EL1.SMEN KVM: arm64: Refactor CPTR trap deactivation KVM: arm64: Refactor exit handlers KVM: arm64: Mark some header functions as inline KVM: arm64: Eagerly switch ZCR_EL{1,2}
arch/arm64/include/asm/kvm_emulate.h | 42 -------- arch/arm64/include/asm/kvm_host.h | 22 +--- arch/arm64/kernel/fpsimd.c | 25 ----- arch/arm64/kvm/arm.c | 8 -- arch/arm64/kvm/fpsimd.c | 100 ++---------------- arch/arm64/kvm/hyp/entry.S | 5 + arch/arm64/kvm/hyp/include/hyp/switch.h | 133 +++++++++++++++++------- arch/arm64/kvm/hyp/nvhe/hyp-main.c | 15 ++- arch/arm64/kvm/hyp/nvhe/switch.c | 91 ++++++++-------- arch/arm64/kvm/hyp/vhe/switch.c | 33 +++--- 10 files changed, 187 insertions(+), 287 deletions(-)