On Wed, Dec 16, 2020 at 08:28:43PM +0800, 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 update the old PTE with a new one if they are different.
>
> (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 update the old PTE with a new
> one if they are different.
>
> The worst case can be like this: vCPU A causes a translation fault with RW
> prot and sets the leaf PTE with RW permissions, and then the next vCPU B
> with RO prot updates the PTE back to RO permissions with break-before-make.
> And the BBM-invalid moment may trigger more unnecessary translation faults,
> then some useless small loops might occur which could lead to vCPU stuck.
>
> To avoid unnecessary update and small loops, add prejudgement in the
> translation fault handler: Skip updating the PTE with break-before-make
> if we are trying to recreate the exact same mapping or only change the
> access permissions. Actually, change of permissions will be handled
> through the relax_perms path next time if necessary.
>
> Signed-off-by: Yanan Wang <[email protected]>
> ---
> arch/arm64/kvm/hyp/pgtable.c | 28 +++++++++++++++++++---------
> 1 file changed, 19 insertions(+), 9 deletions(-)
>
> diff --git a/arch/arm64/kvm/hyp/pgtable.c b/arch/arm64/kvm/hyp/pgtable.c
> index 350f9f810930..8225ced49bad 100644
> --- a/arch/arm64/kvm/hyp/pgtable.c
> +++ b/arch/arm64/kvm/hyp/pgtable.c
> @@ -45,6 +45,10 @@
>
> #define KVM_PTE_LEAF_ATTR_HI_S2_XN BIT(54)
>
> +#define KVM_PTE_LEAF_ATTR_S2_PERMS (KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R | \
> + KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W | \
> + KVM_PTE_LEAF_ATTR_HI_S2_XN)
> +
> struct kvm_pgtable_walk_data {
> struct kvm_pgtable *pgt;
> struct kvm_pgtable_walker *walker;
> @@ -460,7 +464,7 @@ static int stage2_map_set_prot_attr(enum kvm_pgtable_prot prot,
> return 0;
> }
>
> -static bool stage2_map_walker_try_leaf(u64 addr, u64 end, u32 level,
> +static int stage2_map_walker_try_leaf(u64 addr, u64 end, u32 level,
> kvm_pte_t *ptep,
> struct stage2_map_data *data)
> {
> @@ -469,13 +473,18 @@ static bool stage2_map_walker_try_leaf(u64 addr, u64 end, u32 level,
> struct page *page = virt_to_page(ptep);
>
> if (!kvm_block_mapping_supported(addr, end, phys, level))
> - return false;
> + return 1;
It would probably be cleaner to return another error code here, as we
have failed to install a mapping (e.g. E2BIG or perhaps more perversely,
ENOTBLK). Then the caller can decide to install a trable.
Will
On 2021/1/13 23:44, Will Deacon wrote:
> On Wed, Dec 16, 2020 at 08:28:43PM +0800, 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 update the old PTE with a new one if they are different.
>>
>> (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 update the old PTE with a new
>> one if they are different.
>>
>> The worst case can be like this: vCPU A causes a translation fault with RW
>> prot and sets the leaf PTE with RW permissions, and then the next vCPU B
>> with RO prot updates the PTE back to RO permissions with break-before-make.
>> And the BBM-invalid moment may trigger more unnecessary translation faults,
>> then some useless small loops might occur which could lead to vCPU stuck.
>>
>> To avoid unnecessary update and small loops, add prejudgement in the
>> translation fault handler: Skip updating the PTE with break-before-make
>> if we are trying to recreate the exact same mapping or only change the
>> access permissions. Actually, change of permissions will be handled
>> through the relax_perms path next time if necessary.
>>
>> Signed-off-by: Yanan Wang <[email protected]>
>> ---
>> arch/arm64/kvm/hyp/pgtable.c | 28 +++++++++++++++++++---------
>> 1 file changed, 19 insertions(+), 9 deletions(-)
>>
>> diff --git a/arch/arm64/kvm/hyp/pgtable.c b/arch/arm64/kvm/hyp/pgtable.c
>> index 350f9f810930..8225ced49bad 100644
>> --- a/arch/arm64/kvm/hyp/pgtable.c
>> +++ b/arch/arm64/kvm/hyp/pgtable.c
>> @@ -45,6 +45,10 @@
>>
>> #define KVM_PTE_LEAF_ATTR_HI_S2_XN BIT(54)
>>
>> +#define KVM_PTE_LEAF_ATTR_S2_PERMS (KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R | \
>> + KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W | \
>> + KVM_PTE_LEAF_ATTR_HI_S2_XN)
>> +
>> struct kvm_pgtable_walk_data {
>> struct kvm_pgtable *pgt;
>> struct kvm_pgtable_walker *walker;
>> @@ -460,7 +464,7 @@ static int stage2_map_set_prot_attr(enum kvm_pgtable_prot prot,
>> return 0;
>> }
>>
>> -static bool stage2_map_walker_try_leaf(u64 addr, u64 end, u32 level,
>> +static int stage2_map_walker_try_leaf(u64 addr, u64 end, u32 level,
>> kvm_pte_t *ptep,
>> struct stage2_map_data *data)
>> {
>> @@ -469,13 +473,18 @@ static bool stage2_map_walker_try_leaf(u64 addr, u64 end, u32 level,
>> struct page *page = virt_to_page(ptep);
>>
>> if (!kvm_block_mapping_supported(addr, end, phys, level))
>> - return false;
>> + return 1;
> It would probably be cleaner to return another error code here, as we
> have failed to install a mapping (e.g. E2BIG or perhaps more perversely,
> ENOTBLK). Then the caller can decide to install a trable.
>
Ok, I will make some adjustment for this and post the v3 soon after.
Thanks,
Yanan.