On 12/12/2019 14:46, Mark Rutland wrote:
On Thu, Dec 12, 2019 at 03:44:23PM +0530, Anshuman Khandual wrote:
+#define ID_ISAR6_JSCVT_SHIFT 0
+#define ID_ISAR6_DP_SHIFT 4
+#define ID_ISAR6_FHM_SHIFT 8
+#define ID_ISAR6_SB_SHIFT 12
+#define ID_ISAR6_SPECRES_SHIFT 16
+#define ID_ISAR6_BF16_SHIFT 20
+#define ID_ISAR6_I8MM_SHIFT 24
@@ -399,6 +399,7 @@ static const struct __ftr_reg_entry {
ARM64_FTR_REG(SYS_ID_ISAR4_EL1, ftr_generic_32bits),
ARM64_FTR_REG(SYS_ID_ISAR5_EL1, ftr_id_isar5),
ARM64_FTR_REG(SYS_ID_MMFR4_EL1, ftr_id_mmfr4),
+ ARM64_FTR_REG(SYS_ID_ISAR6_EL1, ftr_generic_32bits),
Using ftr_generic_32bits exposes the lowest-common-denominator for all
4-bit fields in the register, and I don't think that's the right thing
to do here, because:
* We have no idea what ID_ISAR6 bits [31:28] may mean in future.
* AFAICT, the instructions described by ID_ISAR6.SPECRES (from the
ARMv8.0-PredInv extension) operate on the local PE and are not
broadcast. To make those work as a guest expects, the host will need
to do additional things (e.g. to preserve that illusion when a vCPU is
migrated from one pCPU to another and back).
Given that, think we should add an explicit ftr_id_isar6 which only
exposes the fields that we're certain are safe to expose to a guest
(i.e. without SPECRES).
Agree. Thanks for pointing this out. I recommended the usage of
generic_32bits table without actually looking at the feature
definitions.
Anshuman,
Sorry about this.
Cheers
Suzuki
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
