| Message ID | 20180504151924.13696-1-james.morse@arm.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 04/05/18 16:19, James Morse wrote: > Proxying the cpuif accesses at EL2 makes use of vcpu_data_guest_to_host > and co, which check the endianness, which call into vcpu_read_sys_reg... > which isn't mapped at EL2 (it was inlined before, and got moved OoL > with the VHE optimizations). > > The result is of course a nice panic. Let's add some specialized > cruft to keep the broken platforms that require this hack alive. > > But, this code used vcpu_data_guest_to_host(), which expected us to > write the value to host memory, instead we have trapped the guest's > read or write to an mmio-device, and are about to replay it using the > host's readl()/writel() which also perform swabbing based on the host > endianness. This goes wrong when both host and guest are big-endian, > as readl()/writel() will undo the guest's swabbing, causing the > big-endian value to be written to device-memory. > > What needs doing? > A big-endian guest will have pre-swabbed data before storing, undo this. > If its necessary for the host, writel() will re-swab it. > > For a read a big-endian guest expects to swab the data after the load. > The hosts's readl() will correct for host endianness, giving us the > device-memory's value in the register. For a big-endian guest, swab it > as if we'd only done the load. > > For a little-endian guest, nothing needs doing as readl()/writel() leave > the correct device-memory value in registers. > > Tested on Juno with that rarest of things: a big-endian 64K host. <song> The Thing That Should Not Be </song> > Based on a patch from Marc Zyngier. > > Reported-by: Suzuki K Poulose <suzuki.poulose@arm.com> > Fixes: bf8feb39642b ("arm64: KVM: vgic-v2: Add GICV access from HYP") > CC: Marc Zyngier <marc.zyngier@arm.com> > Signed-off-by: James Morse <james.morse@arm.com> Awesome, thanks a lot for picking this up and fixing it the right way! I'll queue it right away. Thanks, M.
diff --git a/arch/arm64/kvm/hyp/vgic-v2-cpuif-proxy.c b/arch/arm64/kvm/hyp/vgic-v2-cpuif-proxy.c index 86801b6055d6..39be799d0417 100644 --- a/arch/arm64/kvm/hyp/vgic-v2-cpuif-proxy.c +++ b/arch/arm64/kvm/hyp/vgic-v2-cpuif-proxy.c @@ -18,11 +18,20 @@ #include <linux/compiler.h> #include <linux/irqchip/arm-gic.h> #include <linux/kvm_host.h> +#include <linux/swab.h> #include <asm/kvm_emulate.h> #include <asm/kvm_hyp.h> #include <asm/kvm_mmu.h> +static bool __hyp_text __is_be(struct kvm_vcpu *vcpu) +{ + if (vcpu_mode_is_32bit(vcpu)) + return !!(read_sysreg_el2(spsr) & COMPAT_PSR_E_BIT); + + return !!(read_sysreg(SCTLR_EL1) & SCTLR_ELx_EE); +} + /* * __vgic_v2_perform_cpuif_access -- perform a GICV access on behalf of the * guest. @@ -64,14 +73,19 @@ int __hyp_text __vgic_v2_perform_cpuif_access(struct kvm_vcpu *vcpu) addr += fault_ipa - vgic->vgic_cpu_base; if (kvm_vcpu_dabt_iswrite(vcpu)) { - u32 data = vcpu_data_guest_to_host(vcpu, - vcpu_get_reg(vcpu, rd), - sizeof(u32)); + u32 data = vcpu_get_reg(vcpu, rd); + if (__is_be(vcpu)) { + /* guest pre-swabbed data, undo this for writel() */ + data = swab32(data); + } writel_relaxed(data, addr); } else { u32 data = readl_relaxed(addr); - vcpu_set_reg(vcpu, rd, vcpu_data_host_to_guest(vcpu, data, - sizeof(u32))); + if (__is_be(vcpu)) { + /* guest expects swabbed data */ + data = swab32(data); + } + vcpu_set_reg(vcpu, rd, data); } return 1;
Proxying the cpuif accesses at EL2 makes use of vcpu_data_guest_to_host and co, which check the endianness, which call into vcpu_read_sys_reg... which isn't mapped at EL2 (it was inlined before, and got moved OoL with the VHE optimizations). The result is of course a nice panic. Let's add some specialized cruft to keep the broken platforms that require this hack alive. But, this code used vcpu_data_guest_to_host(), which expected us to write the value to host memory, instead we have trapped the guest's read or write to an mmio-device, and are about to replay it using the host's readl()/writel() which also perform swabbing based on the host endianness. This goes wrong when both host and guest are big-endian, as readl()/writel() will undo the guest's swabbing, causing the big-endian value to be written to device-memory. What needs doing? A big-endian guest will have pre-swabbed data before storing, undo this. If its necessary for the host, writel() will re-swab it. For a read a big-endian guest expects to swab the data after the load. The hosts's readl() will correct for host endianness, giving us the device-memory's value in the register. For a big-endian guest, swab it as if we'd only done the load. For a little-endian guest, nothing needs doing as readl()/writel() leave the correct device-memory value in registers. Tested on Juno with that rarest of things: a big-endian 64K host. Based on a patch from Marc Zyngier. Reported-by: Suzuki K Poulose <suzuki.poulose@arm.com> Fixes: bf8feb39642b ("arm64: KVM: vgic-v2: Add GICV access from HYP") CC: Marc Zyngier <marc.zyngier@arm.com> Signed-off-by: James Morse <james.morse@arm.com> --- This patch doesn't apply before 8a43a2b34b7d ("KVM: arm/arm64: Move arm64-only vgic-v2-sr.c file to arm64"), but the backport is straightforward. arch/arm64/kvm/hyp/vgic-v2-cpuif-proxy.c | 24 +++++++++++++++++++----- 1 file changed, 19 insertions(+), 5 deletions(-)