Hi Yanan,Yes, this is exactly what I want to describe.
On 2020-12-11 08:01, Yanan Wang wrote:
In dirty-logging, or dirty-logging-stopped time, even normal running
time of a guest configed with huge mappings and numbers of vCPUs,
translation faults by different vCPUs on the same GPA could occur
successively almost at the same time. There are two reasons for it.
(1) If there are some vCPUs accessing the same GPA at the same time
and the leaf PTE is not set yet, then they will all cause translation
faults and the first vCPU holding mmu_lock will set valid leaf PTE,
and the others will later choose to update the leaf PTE or not.
(2) When changing a leaf entry or a table entry with break-before-make,
if there are some vCPUs accessing the same GPA just catch the moment
when the target PTE is set invalid in a BBM procedure coincidentally,
they will all cause translation faults and will later choose to update
the leaf PTE or not.
The worst case can be like this: some vCPUs cause translation faults
on the same GPA with different prots, they will fight each other by
changing back access permissions of the PTE with break-before-make.
And the BBM-invalid moment might trigger more unnecessary translation
faults. As a result, some useless small loops will occur, which could
lead to vCPU stuck.
To avoid unnecessary update and small loops, add prejudgement in the
translation fault handler: Skip updating the valid leaf PTE if we are
trying to recreate exactly the same mapping or to reduce access
permissions only(such as RW-->RO). And update the valid leaf PTE without
break-before-make if we are trying to add more permissions only.
I'm a bit perplexed with this: why are you skipping the update if the
permissions need to be reduced? Even more, how can we reduce the
permissions from a vCPU fault? I can't really think of a scenario where
that happens.
Or are you describing a case where two vcpus fault simultaneously with
conflicting permissions:
- Both vcpus fault on translation fault
- vcpu A wants W access
- vpcu B wants R access
and 'A' gets in first, set the permissions to RW (because R is
implicitly added to W), followed by 'B' which downgrades it to RO?
If that's what you are describing, then I agree we could do better.
Signed-off-by: Yanan Wang <wangyanan55@xxxxxxxxxx>
---
arch/arm64/kvm/hyp/pgtable.c | 73 +++++++++++++++++++++++++-----------
1 file changed, 52 insertions(+), 21 deletions(-)
diff --git a/arch/arm64/kvm/hyp/pgtable.c b/arch/arm64/kvm/hyp/pgtable.c
index 23a01dfcb27a..f8b3248cef1c 100644
--- a/arch/arm64/kvm/hyp/pgtable.c
+++ b/arch/arm64/kvm/hyp/pgtable.c
@@ -45,6 +45,8 @@
#define KVM_PTE_LEAF_ATTR_HI_S2_XN BIT(54)
+#define KVM_PTE_LEAF_ATTR_PERMS (GENMASK(7, 6) | BIT(54))
+
struct kvm_pgtable_walk_data {
struct kvm_pgtable *pgt;
struct kvm_pgtable_walker *walker;
@@ -170,10 +172,9 @@ static void kvm_set_table_pte(kvm_pte_t *ptep,
kvm_pte_t *childp)
smp_store_release(ptep, pte);
}
-static bool kvm_set_valid_leaf_pte(kvm_pte_t *ptep, u64 pa, kvm_pte_t attr,
- u32 level)
+static kvm_pte_t kvm_init_valid_leaf_pte(u64 pa, kvm_pte_t attr, u32 level)
{
- kvm_pte_t old = *ptep, pte = kvm_phys_to_pte(pa);
+ kvm_pte_t pte = kvm_phys_to_pte(pa);
u64 type = (level == KVM_PGTABLE_MAX_LEVELS - 1) ? KVM_PTE_TYPE_PAGE :
KVM_PTE_TYPE_BLOCK;
@@ -181,12 +182,7 @@ static bool kvm_set_valid_leaf_pte(kvm_pte_t
*ptep, u64 pa, kvm_pte_t attr,
pte |= FIELD_PREP(KVM_PTE_TYPE, type);
pte |= KVM_PTE_VALID;
- /* Tolerate KVM recreating the exact same mapping. */
- if (kvm_pte_valid(old))
- return old == pte;
-
- smp_store_release(ptep, pte);
- return true;
+ return pte;
}
static int kvm_pgtable_visitor_cb(struct kvm_pgtable_walk_data *data, u64 addr,
@@ -341,12 +337,17 @@ static int hyp_map_set_prot_attr(enum
kvm_pgtable_prot prot,
static bool hyp_map_walker_try_leaf(u64 addr, u64 end, u32 level,
kvm_pte_t *ptep, struct hyp_map_data *data)
{
+ kvm_pte_t new, old = *ptep;
u64 granule = kvm_granule_size(level), phys = data->phys;
if (!kvm_block_mapping_supported(addr, end, phys, level))
return false;
- WARN_ON(!kvm_set_valid_leaf_pte(ptep, phys, data->attr, level));
+ /* Tolerate KVM recreating the exact same mapping. */
+ new = kvm_init_valid_leaf_pte(phys, data->attr, level);
+ if (old != new && !WARN_ON(kvm_pte_valid(old)))
+ smp_store_release(ptep, new);
+
data->phys += granule;
return true;
}
@@ -461,25 +462,56 @@ static int stage2_map_set_prot_attr(enum
kvm_pgtable_prot prot,
return 0;
}
+static bool stage2_set_valid_leaf_pte_pre(u64 addr, u32 level,
+ kvm_pte_t *ptep, kvm_pte_t new,
+ struct stage2_map_data *data)
+{
+ kvm_pte_t old = *ptep, old_attr, new_attr;
+
+ if ((old ^ new) & (~KVM_PTE_LEAF_ATTR_PERMS))
+ return false;
+
+ /*
+ * Skip updating if we are trying to recreate exactly the same mapping
+ * or to reduce the access permissions only. And update the valid leaf
+ * PTE without break-before-make if we are trying to add more access
+ * permissions only.
+ */
+ old_attr = (old & KVM_PTE_LEAF_ATTR_PERMS) ^ KVM_PTE_LEAF_ATTR_HI_S2_XN;
+ new_attr = (new & KVM_PTE_LEAF_ATTR_PERMS) ^ KVM_PTE_LEAF_ATTR_HI_S2_XN;
+ if (new_attr <= old_attr)
+ return true;
+
+ WRITE_ONCE(*ptep, new);
+ kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, data->mmu, addr, level);
I think what bothers me the most here is that we are turning a mapping into
a permission update, which makes the code really hard to read, and mixes
two things that were so far separate.
I wonder whether we should instead abort the update and simply take the fault
again, if we ever need to do it.
Thanks,
M.