diff mbox series

[5/9] KVM: arm: move has_run_once after the map_resources

Message ID 20201212185010.26579-6-eric.auger@redhat.com (mailing list archive)
State New, archived
Headers show
Series KVM/ARM: Some vgic fixes and init sequence KVM selftests | expand

Commit Message

Eric Auger Dec. 12, 2020, 6:50 p.m. UTC
has_run_once is set to true at the beginning of
kvm_vcpu_first_run_init(). This generally is not an issue
except when exercising the code with KVM selftests. Indeed,
if kvm_vgic_map_resources() fails due to erroneous user settings,
has_run_once is set and this prevents from continuing
executing the test. This patch moves the assignment after the
kvm_vgic_map_resources().

Signed-off-by: Eric Auger <eric.auger@redhat.com>
---
 arch/arm64/kvm/arm.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

Comments

Alexandru Elisei Jan. 12, 2021, 2:55 p.m. UTC | #1
Hi Eric,

On 12/12/20 6:50 PM, Eric Auger wrote:
> has_run_once is set to true at the beginning of
> kvm_vcpu_first_run_init(). This generally is not an issue
> except when exercising the code with KVM selftests. Indeed,
> if kvm_vgic_map_resources() fails due to erroneous user settings,
> has_run_once is set and this prevents from continuing
> executing the test. This patch moves the assignment after the
> kvm_vgic_map_resources().
>
> Signed-off-by: Eric Auger <eric.auger@redhat.com>
> ---
>  arch/arm64/kvm/arm.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
> index c0ffb019ca8b..331fae6bff31 100644
> --- a/arch/arm64/kvm/arm.c
> +++ b/arch/arm64/kvm/arm.c
> @@ -540,8 +540,6 @@ static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
>  	if (!kvm_arm_vcpu_is_finalized(vcpu))
>  		return -EPERM;
>  
> -	vcpu->arch.has_run_once = true;
> -
>  	if (likely(irqchip_in_kernel(kvm))) {
>  		/*
>  		 * Map the VGIC hardware resources before running a vcpu the
> @@ -560,6 +558,8 @@ static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
>  		static_branch_inc(&userspace_irqchip_in_use);
>  	}
>  
> +	vcpu->arch.has_run_once = true;

I have a few concerns regarding this:

1. Moving has_run_once = true here seems very arbitrary to me - kvm_timer_enable()
and kvm_arm_pmu_v3_enable(), below it, can both fail because of erroneous user
values. If there's a reason why the assignment cannot be moved at the end of the
function, I think it should be clearly stated in a comment for the people who
might be tempted to write similar tests for the timer or pmu.

2. There are many ways that kvm_vgic_map_resources() can fail, other than
incorrect user settings. I started digging into how
kvm_vgic_map_resources()->vgic_v2_map_resources() can fail for a VGIC V2 and this
is what I managed to find before I gave up:

* vgic_init() can fail in:
    - kvm_vgic_dist_init()
    - vgic_v3_init()
    - kvm_vgic_setup_default_irq_routing()
* vgic_register_dist_iodev() can fail in:
    - vgic_v3_init_dist_iodev()
    - kvm_io_bus_register_dev()(*)
* kvm_phys_addr_ioremap() can fail in:
    - kvm_mmu_topup_memory_cache()
    - kvm_pgtable_stage2_map()

So if any of the functions below fail, are we 100% sure it is safe to allow the
user to execute kvm_vgic_map_resources() again?

(*) It looks to me like kvm_io_bus_register_dev() doesn't take into account a
caller that tries to register the same device address range and it will create
another identical range. Is this intentional? Is it a bug that should be fixed? Or
am I misunderstanding the function?

Thanks,
Alex
> +
>  	ret = kvm_timer_enable(vcpu);
>  	if (ret)
>  		return ret;
Eric Auger Jan. 14, 2021, 10:02 a.m. UTC | #2
Hi Alexandru,

On 1/12/21 3:55 PM, Alexandru Elisei wrote:
> Hi Eric,
> 
> On 12/12/20 6:50 PM, Eric Auger wrote:
>> has_run_once is set to true at the beginning of
>> kvm_vcpu_first_run_init(). This generally is not an issue
>> except when exercising the code with KVM selftests. Indeed,
>> if kvm_vgic_map_resources() fails due to erroneous user settings,
>> has_run_once is set and this prevents from continuing
>> executing the test. This patch moves the assignment after the
>> kvm_vgic_map_resources().
>>
>> Signed-off-by: Eric Auger <eric.auger@redhat.com>
>> ---
>>  arch/arm64/kvm/arm.c | 4 ++--
>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
>> index c0ffb019ca8b..331fae6bff31 100644
>> --- a/arch/arm64/kvm/arm.c
>> +++ b/arch/arm64/kvm/arm.c
>> @@ -540,8 +540,6 @@ static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
>>  	if (!kvm_arm_vcpu_is_finalized(vcpu))
>>  		return -EPERM;
>>  
>> -	vcpu->arch.has_run_once = true;
>> -
>>  	if (likely(irqchip_in_kernel(kvm))) {
>>  		/*
>>  		 * Map the VGIC hardware resources before running a vcpu the
>> @@ -560,6 +558,8 @@ static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
>>  		static_branch_inc(&userspace_irqchip_in_use);
>>  	}
>>  
>> +	vcpu->arch.has_run_once = true;
> 
> I have a few concerns regarding this:
> 
> 1. Moving has_run_once = true here seems very arbitrary to me - kvm_timer_enable()
> and kvm_arm_pmu_v3_enable(), below it, can both fail because of erroneous user
> values. If there's a reason why the assignment cannot be moved at the end of the
> function, I think it should be clearly stated in a comment for the people who
> might be tempted to write similar tests for the timer or pmu.

Setting has_run_once = true at the entry of the function looks to me
even more arbitrary. I agree with you that eventually has_run_once may
be moved at the very end but maybe this can be done later once timer,
pmu tests haven ben written
> 
> 2. There are many ways that kvm_vgic_map_resources() can fail, other than
> incorrect user settings. I started digging into how
> kvm_vgic_map_resources()->vgic_v2_map_resources() can fail for a VGIC V2 and this
> is what I managed to find before I gave up:
> 
> * vgic_init() can fail in:
>     - kvm_vgic_dist_init()
>     - vgic_v3_init()
>     - kvm_vgic_setup_default_irq_routing()
> * vgic_register_dist_iodev() can fail in:
>     - vgic_v3_init_dist_iodev()
>     - kvm_io_bus_register_dev()(*)
> * kvm_phys_addr_ioremap() can fail in:
>     - kvm_mmu_topup_memory_cache()
>     - kvm_pgtable_stage2_map()

I changed the commit msg so that "incorrect user settings" sounds as an
example.
> 
> So if any of the functions below fail, are we 100% sure it is safe to allow the
> user to execute kvm_vgic_map_resources() again?

I think additional tests will confirm this. However at the moment,
moving the assignment, which does not look wrong to me, allows to
greatly simplify the tests so I would tend to say that it is worth.
> 
> (*) It looks to me like kvm_io_bus_register_dev() doesn't take into account a
> caller that tries to register the same device address range and it will create
> another identical range. Is this intentional? Is it a bug that should be fixed? Or
> am I misunderstanding the function?

doesn't kvm_io_bus_cmp() do the check?

Thanks

Eric
> 
> Thanks,
> Alex
>> +
>>  	ret = kvm_timer_enable(vcpu);
>>  	if (ret)
>>  		return ret;
>
Alexandru Elisei Jan. 20, 2021, 3:56 p.m. UTC | #3
Hi Eric,

On 1/14/21 10:02 AM, Auger Eric wrote:
> Hi Alexandru,
>
> On 1/12/21 3:55 PM, Alexandru Elisei wrote:
>> Hi Eric,
>>
>> On 12/12/20 6:50 PM, Eric Auger wrote:
>>> has_run_once is set to true at the beginning of
>>> kvm_vcpu_first_run_init(). This generally is not an issue
>>> except when exercising the code with KVM selftests. Indeed,
>>> if kvm_vgic_map_resources() fails due to erroneous user settings,
>>> has_run_once is set and this prevents from continuing
>>> executing the test. This patch moves the assignment after the
>>> kvm_vgic_map_resources().
>>>
>>> Signed-off-by: Eric Auger <eric.auger@redhat.com>
>>> ---
>>>  arch/arm64/kvm/arm.c | 4 ++--
>>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
>>> index c0ffb019ca8b..331fae6bff31 100644
>>> --- a/arch/arm64/kvm/arm.c
>>> +++ b/arch/arm64/kvm/arm.c
>>> @@ -540,8 +540,6 @@ static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
>>>  	if (!kvm_arm_vcpu_is_finalized(vcpu))
>>>  		return -EPERM;
>>>  
>>> -	vcpu->arch.has_run_once = true;
>>> -
>>>  	if (likely(irqchip_in_kernel(kvm))) {
>>>  		/*
>>>  		 * Map the VGIC hardware resources before running a vcpu the
>>> @@ -560,6 +558,8 @@ static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
>>>  		static_branch_inc(&userspace_irqchip_in_use);
>>>  	}
>>>  
>>> +	vcpu->arch.has_run_once = true;
>> I have a few concerns regarding this:
>>
>> 1. Moving has_run_once = true here seems very arbitrary to me - kvm_timer_enable()
>> and kvm_arm_pmu_v3_enable(), below it, can both fail because of erroneous user
>> values. If there's a reason why the assignment cannot be moved at the end of the
>> function, I think it should be clearly stated in a comment for the people who
>> might be tempted to write similar tests for the timer or pmu.
> Setting has_run_once = true at the entry of the function looks to me
> even more arbitrary. I agree with you that eventually has_run_once may

Or it could be it's there to prevent the user from calling
kvm_vgic_map_resources() a second time after it failed. This is what I'm concerned
about and I think deserves more investigation.

Thanks,
Alex
> be moved at the very end but maybe this can be done later once timer,
> pmu tests haven ben written
>> 2. There are many ways that kvm_vgic_map_resources() can fail, other than
>> incorrect user settings. I started digging into how
>> kvm_vgic_map_resources()->vgic_v2_map_resources() can fail for a VGIC V2 and this
>> is what I managed to find before I gave up:
>>
>> * vgic_init() can fail in:
>>     - kvm_vgic_dist_init()
>>     - vgic_v3_init()
>>     - kvm_vgic_setup_default_irq_routing()
>> * vgic_register_dist_iodev() can fail in:
>>     - vgic_v3_init_dist_iodev()
>>     - kvm_io_bus_register_dev()(*)
>> * kvm_phys_addr_ioremap() can fail in:
>>     - kvm_mmu_topup_memory_cache()
>>     - kvm_pgtable_stage2_map()
> I changed the commit msg so that "incorrect user settings" sounds as an
> example.
>> So if any of the functions below fail, are we 100% sure it is safe to allow the
>> user to execute kvm_vgic_map_resources() again?
> I think additional tests will confirm this. However at the moment,
> moving the assignment, which does not look wrong to me, allows to
> greatly simplify the tests so I would tend to say that it is worth.
>> (*) It looks to me like kvm_io_bus_register_dev() doesn't take into account a
>> caller that tries to register the same device address range and it will create
>> another identical range. Is this intentional? Is it a bug that should be fixed? Or
>> am I misunderstanding the function?
> doesn't kvm_io_bus_cmp() do the check?
>
> Thanks
>
> Eric
>> Thanks,
>> Alex
>>> +
>>>  	ret = kvm_timer_enable(vcpu);
>>>  	if (ret)
>>>  		return ret;
Eric Auger March 12, 2021, 5:27 p.m. UTC | #4
Hi Alexandru,

On 1/20/21 4:56 PM, Alexandru Elisei wrote:
> Hi Eric,
> 
> On 1/14/21 10:02 AM, Auger Eric wrote:
>> Hi Alexandru,
>>
>> On 1/12/21 3:55 PM, Alexandru Elisei wrote:
>>> Hi Eric,
>>>
>>> On 12/12/20 6:50 PM, Eric Auger wrote:
>>>> has_run_once is set to true at the beginning of
>>>> kvm_vcpu_first_run_init(). This generally is not an issue
>>>> except when exercising the code with KVM selftests. Indeed,
>>>> if kvm_vgic_map_resources() fails due to erroneous user settings,
>>>> has_run_once is set and this prevents from continuing
>>>> executing the test. This patch moves the assignment after the
>>>> kvm_vgic_map_resources().
>>>>
>>>> Signed-off-by: Eric Auger <eric.auger@redhat.com>
>>>> ---
>>>>  arch/arm64/kvm/arm.c | 4 ++--
>>>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
>>>> index c0ffb019ca8b..331fae6bff31 100644
>>>> --- a/arch/arm64/kvm/arm.c
>>>> +++ b/arch/arm64/kvm/arm.c
>>>> @@ -540,8 +540,6 @@ static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
>>>>  	if (!kvm_arm_vcpu_is_finalized(vcpu))
>>>>  		return -EPERM;
>>>>  
>>>> -	vcpu->arch.has_run_once = true;
>>>> -
>>>>  	if (likely(irqchip_in_kernel(kvm))) {
>>>>  		/*
>>>>  		 * Map the VGIC hardware resources before running a vcpu the
>>>> @@ -560,6 +558,8 @@ static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
>>>>  		static_branch_inc(&userspace_irqchip_in_use);
>>>>  	}
>>>>  
>>>> +	vcpu->arch.has_run_once = true;
>>> I have a few concerns regarding this:
>>>
>>> 1. Moving has_run_once = true here seems very arbitrary to me - kvm_timer_enable()
>>> and kvm_arm_pmu_v3_enable(), below it, can both fail because of erroneous user
>>> values. If there's a reason why the assignment cannot be moved at the end of the
>>> function, I think it should be clearly stated in a comment for the people who
>>> might be tempted to write similar tests for the timer or pmu.
>> Setting has_run_once = true at the entry of the function looks to me
>> even more arbitrary. I agree with you that eventually has_run_once may
> 
> Or it could be it's there to prevent the user from calling
> kvm_vgic_map_resources() a second time after it failed. This is what I'm concerned
> about and I think deserves more investigation.

I have reworked my kvm selftests to live without that change.

Thanks

Eric
> 
> Thanks,
> Alex
>> be moved at the very end but maybe this can be done later once timer,
>> pmu tests haven ben written
>>> 2. There are many ways that kvm_vgic_map_resources() can fail, other than
>>> incorrect user settings. I started digging into how
>>> kvm_vgic_map_resources()->vgic_v2_map_resources() can fail for a VGIC V2 and this
>>> is what I managed to find before I gave up:
>>>
>>> * vgic_init() can fail in:
>>>     - kvm_vgic_dist_init()
>>>     - vgic_v3_init()
>>>     - kvm_vgic_setup_default_irq_routing()
>>> * vgic_register_dist_iodev() can fail in:
>>>     - vgic_v3_init_dist_iodev()
>>>     - kvm_io_bus_register_dev()(*)
>>> * kvm_phys_addr_ioremap() can fail in:
>>>     - kvm_mmu_topup_memory_cache()
>>>     - kvm_pgtable_stage2_map()
>> I changed the commit msg so that "incorrect user settings" sounds as an
>> example.
>>> So if any of the functions below fail, are we 100% sure it is safe to allow the
>>> user to execute kvm_vgic_map_resources() again?
>> I think additional tests will confirm this. However at the moment,
>> moving the assignment, which does not look wrong to me, allows to
>> greatly simplify the tests so I would tend to say that it is worth.
>>> (*) It looks to me like kvm_io_bus_register_dev() doesn't take into account a
>>> caller that tries to register the same device address range and it will create
>>> another identical range. Is this intentional? Is it a bug that should be fixed? Or
>>> am I misunderstanding the function?
>> doesn't kvm_io_bus_cmp() do the check?
>>
>> Thanks
>>
>> Eric
>>> Thanks,
>>> Alex
>>>> +
>>>>  	ret = kvm_timer_enable(vcpu);
>>>>  	if (ret)
>>>>  		return ret;
>
diff mbox series

Patch

diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
index c0ffb019ca8b..331fae6bff31 100644
--- a/arch/arm64/kvm/arm.c
+++ b/arch/arm64/kvm/arm.c
@@ -540,8 +540,6 @@  static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
 	if (!kvm_arm_vcpu_is_finalized(vcpu))
 		return -EPERM;
 
-	vcpu->arch.has_run_once = true;
-
 	if (likely(irqchip_in_kernel(kvm))) {
 		/*
 		 * Map the VGIC hardware resources before running a vcpu the
@@ -560,6 +558,8 @@  static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
 		static_branch_inc(&userspace_irqchip_in_use);
 	}
 
+	vcpu->arch.has_run_once = true;
+
 	ret = kvm_timer_enable(vcpu);
 	if (ret)
 		return ret;