[PATCH 13/16] KVM: arm64: nv: Invalidate TLBs based on shadow S2 TTL-like information

[Marc Zyngier]

In order to be able to make S2 TLB invalidations more performant on NV,
let's use a scheme derived from the FEAT_TTL extension.

If bits [56:55] in the leaf descriptor translating the address in the
corresponding shadow S2 are non-zero, they indicate a level which can
be used as an invalidation range. This allows further reduction of the
systematic over-invalidation that takes place otherwise.

Signed-off-by: Marc Zyngier <maz at kernel.org>
---
 arch/arm64/kvm/nested.c | 83 ++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 82 insertions(+), 1 deletion(-)

diff --git a/arch/arm64/kvm/nested.c b/arch/arm64/kvm/nested.c
index 5abdf82bb9a4..8163f7308b82 100644
--- a/arch/arm64/kvm/nested.c
+++ b/arch/arm64/kvm/nested.c
@@ -4,6 +4,7 @@
  * Author: Jintack Lim <jintack.lim at linaro.org>
  */
 
+#include <linux/bitfield.h>
 #include <linux/kvm.h>
 #include <linux/kvm_host.h>
 
@@ -405,12 +406,92 @@ static unsigned int ttl_to_size(u8 ttl)
 	return max_size;
 }
 
+/*
+ * Compute the equivalent of the TTL field by parsing the shadow PT.  The
+ * granule size is extracted from the cached VTCR_EL2.TG0 while the level is
+ * retrieved from first entry carrying the level as a tag.
+ */
+static u8 get_guest_mapping_ttl(struct kvm_s2_mmu *mmu, u64 addr)
+{
+	u64 tmp, sz = 0, vtcr = mmu->tlb_vtcr;
+	kvm_pte_t pte;
+	u8 ttl, level;
+
+	switch (vtcr & VTCR_EL2_TG0_MASK) {
+	case VTCR_EL2_TG0_4K:
+		ttl = (TLBI_TTL_TG_4K << 2);
+		break;
+	case VTCR_EL2_TG0_16K:
+		ttl = (TLBI_TTL_TG_16K << 2);
+		break;
+	case VTCR_EL2_TG0_64K:
+	default:	    /* IMPDEF: treat any other value as 64k */
+		ttl = (TLBI_TTL_TG_64K << 2);
+		break;
+	}
+
+	tmp = addr;
+
+again:
+	/* Iteratively compute the block sizes for a particular granule size */
+	switch (vtcr & VTCR_EL2_TG0_MASK) {
+	case VTCR_EL2_TG0_4K:
+		if	(sz < SZ_4K)	sz = SZ_4K;
+		else if (sz < SZ_2M)	sz = SZ_2M;
+		else if (sz < SZ_1G)	sz = SZ_1G;
+		else			sz = 0;
+		break;
+	case VTCR_EL2_TG0_16K:
+		if	(sz < SZ_16K)	sz = SZ_16K;
+		else if (sz < SZ_32M)	sz = SZ_32M;
+		else			sz = 0;
+		break;
+	case VTCR_EL2_TG0_64K:
+	default:	    /* IMPDEF: treat any other value as 64k */
+		if	(sz < SZ_64K)	sz = SZ_64K;
+		else if (sz < SZ_512M)	sz = SZ_512M;
+		else			sz = 0;
+		break;
+	}
+
+	if (sz == 0)
+		return 0;
+
+	tmp &= ~(sz - 1);
+	if (kvm_pgtable_get_leaf(mmu->pgt, tmp, &pte, NULL))
+		goto again;
+	if (!(pte & PTE_VALID))
+		goto again;
+	level = FIELD_GET(KVM_NV_GUEST_MAP_SZ, pte);
+	if (!level)
+		goto again;
+
+	ttl |= level;
+
+	/*
+	 * We now have found some level information in the shadow S2. Check
+	 * that the resulting range is actually including the original IPA.
+	 */
+	sz = ttl_to_size(ttl);
+	if (addr < (tmp + sz))
+		return ttl;
+
+	return 0;
+}
+
 unsigned long compute_tlb_inval_range(struct kvm_s2_mmu *mmu, u64 val)
 {
+	struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
 	unsigned long max_size;
 	u8 ttl;
 
-	ttl = FIELD_GET(GENMASK_ULL(47, 44), val);
+	ttl = FIELD_GET(TLBI_TTL_MASK, val);
+
+	if (!ttl || !kvm_has_feat(kvm, ID_AA64MMFR2_EL1, TTL, IMP)) {
+		/* No TTL, check the shadow S2 for a hint */
+		u64 addr = (val & GENMASK_ULL(35, 0)) << 12;
+		ttl = get_guest_mapping_ttl(mmu, addr);
+	}
 
 	max_size = ttl_to_size(ttl);
 
-- 
2.39.2