Re: [PATCH v6 1/2] arm64: kvm: Save/restore MTE registers

From: Marc Zyngier
Date: Mon Dec 07 2020 - 10:57:05 EST


On 2020-12-07 14:48, Steven Price wrote:
On 03/12/2020 17:07, Marc Zyngier wrote:

diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
index e2ef4c2edf06..b6668ffa04d9 100644
--- a/arch/arm64/include/asm/sysreg.h
+++ b/arch/arm64/include/asm/sysreg.h
@@ -569,7 +569,8 @@
 #define SCTLR_ELx_M    (BIT(0))

 #define SCTLR_ELx_FLAGS    (SCTLR_ELx_M  | SCTLR_ELx_A | SCTLR_ELx_C | \
-             SCTLR_ELx_SA | SCTLR_ELx_I | SCTLR_ELx_IESB)
+             SCTLR_ELx_SA | SCTLR_ELx_I | SCTLR_ELx_IESB | \
+             SCTLR_ELx_ITFSB)

 /* SCTLR_EL2 specific flags. */
 #define SCTLR_EL2_RES1    ((BIT(4))  | (BIT(5))  | (BIT(11)) | (BIT(16)) | \
diff --git a/arch/arm64/kvm/hyp/include/hyp/sysreg-sr.h
b/arch/arm64/kvm/hyp/include/hyp/sysreg-sr.h
index cce43bfe158f..45255ba60152 100644
--- a/arch/arm64/kvm/hyp/include/hyp/sysreg-sr.h
+++ b/arch/arm64/kvm/hyp/include/hyp/sysreg-sr.h
@@ -18,6 +18,11 @@
 static inline void __sysreg_save_common_state(struct kvm_cpu_context *ctxt)
 {
     ctxt_sys_reg(ctxt, MDSCR_EL1)    = read_sysreg(mdscr_el1);
+    if (system_supports_mte()) {

Please move the per-VM predicate to this patch so that it can be used
not to save/restore the MTE registers if we don't need to.

There isn't a valid struct kvm_vcpu or struct kvm here. I know there's
ctx->__hyp_running_vcpu but AFAICT that is only valid with the host
context.

We have per-CPU variables for the host context. If ctx->__hyp_running_vcpu
is non NULL, you know you're on the host.




+        ctxt_sys_reg(ctxt, RGSR_EL1)    = read_sysreg_s(SYS_RGSR_EL1);
+        ctxt_sys_reg(ctxt, GCR_EL1)    = read_sysreg_s(SYS_GCR_EL1);
+        ctxt_sys_reg(ctxt, TFSRE0_EL1)    = read_sysreg_s(SYS_TFSRE0_EL1);
+    }

Overall, I still don't understand how this is going to work once
we have MTE in the kernel. You mentioned having the ability to
create turn off the tag checks at times, but I don't see that
in this patch (and I'm not sure we want that either).

Given that this is now highly unlikely to be merged for v5.11, I'll
rebase onto of the KASAN MTE series and double check exactly what
happens. My thought was that it should be as simple as setting TCO,
but your previous comment about moving the save/restore into assembler
might be wise in case the compiler starts playing with TCO itself.

Indeed.

Thanks,

M.
--
Jazz is not dead. It just smells funny...