lkml.org 
[lkml]   [2017]   [Jul]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH v2] KVM: arm/arm64: Handle hva aging while destroying the vm
On Mon, Jul 17, 2017 at 3:03 PM, Suzuki K Poulose
<Suzuki.Poulose@arm.com> wrote:
> On 16/07/17 20:56, Christoffer Dall wrote:
>>
>> On Fri, Jul 14, 2017 at 05:40:48PM +0100, Suzuki K Poulose wrote:
>>>
>>> On 06/07/17 10:42, Christoffer Dall wrote:
>>>>
>>>> On Thu, Jul 06, 2017 at 10:34:58AM +0100, Suzuki K Poulose wrote:
>>>>>
>>>>> On 06/07/17 08:45, Christoffer Dall wrote:
>>>>>>
>>>>>> On Thu, Jul 06, 2017 at 09:07:49AM +0200, Alexander Graf wrote:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On 05.07.17 10:57, Suzuki K Poulose wrote:
>>>>>>>>
>>>>>>>> Hi Alex,
>>>>>>>>
>>>>>>>> On Wed, Jul 05, 2017 at 08:20:31AM +0200, Alexander Graf wrote:
>>>>>>>>>
>>>>>>>>> The kvm_age_hva callback may be called all the way concurrently
>>>>>>>>> while
>>>>>>>>> kvm_mmu_notifier_release() is running.
>>>>>>>>>
>>>>>>>>> The release function sets kvm->arch.pgd = NULL which the aging
>>>>>>>>> function
>>>>>>>>> however implicitly relies on in stage2_get_pud(). That means they
>>>>>>>>> can
>>>>>>>>> race and the aging function may dereference a NULL pgd pointer.
>>>>>>>>>
>>>>>>>>> This patch adds a check for that case, so that we leave the aging
>>>>>>>>> function silently.
>>>>>>>>>
>>>>>>>>> Cc: stable@vger.kernel.org
>>>>>>>>> Fixes: 293f29363 ("kvm-arm: Unmap shadow pagetables properly")
>>>>>>>>> Signed-off-by: Alexander Graf <agraf@suse.de>
>>>>>>>>>
>>>>>>>>> ---
>>>>>>>>>
>>>>>>>>> v1 -> v2:
>>>>>>>>>
>>>>>>>>> - Fix commit message
>>>>>>>>> - Add Fixes and stable tags
>>>>>>>>> ---
>>>>>>>>> virt/kvm/arm/mmu.c | 4 ++++
>>>>>>>>> 1 file changed, 4 insertions(+)
>>>>>>>>>
>>>>>>>>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
>>>>>>>>> index f2d5b6c..227931f 100644
>>>>>>>>> --- a/virt/kvm/arm/mmu.c
>>>>>>>>> +++ b/virt/kvm/arm/mmu.c
>>>>>>>>> @@ -861,6 +861,10 @@ static pud_t *stage2_get_pud(struct kvm *kvm,
>>>>>>>>> struct kvm_mmu_memory_cache *cache
>>>>>>>>> pgd_t *pgd;
>>>>>>>>> pud_t *pud;
>>>>>>>>> + /* Do we clash with kvm_free_stage2_pgd()? */
>>>>>>>>> + if (!kvm->arch.pgd)
>>>>>>>>> + return NULL;
>>>>>>>>> +
>>>>>>>>
>>>>>>>>
>>>>>>>> I think this check should be moved up in the chain. We call
>>>>>>>> kvm_age_hva(), with
>>>>>>>> the kvm->mmu_lock held and we don't release it till we reach here.
>>>>>>>> So, ideally,
>>>>>>>> if we find the PGD is null when we reach kvm_age_hva(), we could
>>>>>>>> simply return
>>>>>>>> there, like we do for other call backs from the KVM mmu_notifier.
>>>>>>>
>>>>>>>
>>>>>>> That probably works too - I'm not sure which version is more
>>>>>>> consistent as well as more maintainable in the long run. I'll leave
>>>>>>> the call here to Christoffer.
>>>>>>>
>>>>>>
>>>>>> Let's look at the callers to stage2_get_pmd, which is the only caller
>>>>>> of
>>>>>> stage2_get_pud, where the problem was observed:
>>>>>>
>>>>>> user_mem_abort
>>>>>> -> stage2_set_pmd_huge
>>>>>> -> stage2_get_pmd
>>>>>>
>>>>>> user_mem_abort
>>>>>> -> stage2_set_pte
>>>>>> -> stage2_get_pmd
>>>>>>
>>>>>> handle_access_fault
>>>>>> -> stage2_get_pmd
>>>>>>
>>>>>> For the above three functions, pgd cannot ever be NULL, because this
>>>>>> is
>>>>>> running in the context of a VCPU thread, which means the reference on
>>>>>> the VM fd must not reach zero, so no need to call that here.
>>>>>
>>>>>
>>>>> I think there is some problem here. See below for more information.
>>>>>
>>>>>>
>>>>>> kvm_set_spte_handler
>>>>>> -> stage2_set_pte
>>>>>> -> stage2_get_pmd
>>>>>>
>>>>>> This is called from kvm_set_spte_hva, which is one of the MMU
>>>>>> notifiers,
>>>>>> so it can race similarly kvm_age_hva and kvm_test_age_hva, but it
>>>>>> already checks for !kvm->arch.pgd.
>>>>>>
>>>>>> kvm_phys_addr_ioremap
>>>>>> -> stage2_set_pte
>>>>>> -> stage2_get_pmd
>>>>>>
>>>>>> This is called from two places: (1) The VGIC code (as part of
>>>>>> vgic_v2_map_resources) and can only be called in the context of
>>>>>> running
>>>>>> a VCPU, so the pgd cannot be null by virtue of the same argument as
>>>>>> for
>>>>>> user_mem_abort. (2) kvm_arch_prepare_memory_region calls
>>>>>> kvm_phys_addr_ioremap, which is a VM ioctl so similarly, I cannot see
>>>>>> how the VM can be in the middle of being freed while handling ioctls
>>>>>> on
>>>>>> the fd. Therefore, following the same argument, this should be safe
>>>>>> as
>>>>>> well.
>>>>>>
>>>>>> kvm_age_hva_handler and kvm_test_age_hva_handler
>>>>>> -> stage2_get_pmd
>>>>>>
>>>>>> Handled by the patch proposed by Suzuki.
>>>>>>
>>>>>> What does all that tell us? First, it should give us some comfort
>>>>>> that we
>>>>>> don't have more races of this kind. Second, it teels us that there
>>>>>> are
>>>>>> a number of different and not-obvious call paths to stage2_pet_pud,
>>>>>> which could be an argument to simply check the pgd whenever it's
>>>>>> called,
>>>>>> despite the minor runtime overhead. On the other hand, the check
>>>>>> itself
>>>>>> is only valid knowing that we synchronize against kvm_free_stage2_pgd
>>>>>> using the kvm->mmu_lock() and understanding that this only happens
>>>>>> when
>>>>>> mmu notifiers call into the KVM MMU code outside the context of the
>>>>>> VM.
>>>>>>
>>>>>> The last consideration is the winning argument for me to put the check
>>>>>> in kvm_age_hva_handler and kvm_test_age_hva_handler, but I think it's
>>>>>> important that we document why it's only these three high-level
>>>>>> callers
>>>>>> (incl. kvm_set_spte_handler) that need the check, either in the code
>>>>>> or
>>>>>> in the commit message.
>>>>>
>>>>>
>>>>> The only way we end up freeing the stage-2 PGD is via the
>>>>> mmu_notifier_release(),
>>>>> which could be triggered via two different paths.
>>>>>
>>>>> 1) kvm_destroy_vm(), where all the VM resources has been released and
>>>>> the
>>>>> refcount on the KVM instances are dropped, via kvm_put_kvm().
>>>>>
>>>>> kvm_put_kvm()
>>>>> kvm_destroy_vm()
>>>>> mmu_notifier_unregsiter
>>>>> mmu_notifier_ops->release()
>>>>> kvm_arch_flush_shadow_all
>>>>> kvm_free_stage2_pgd -> free the page table with the mmu_lock
>>>>> held
>>>>> occasionally releasing it to avoid
>>>>> contention.
>>>>> or
>>>>>
>>>>> 2) do_signal -> get_signal -> do_group_exit - >
>>>>> do_exit
>>>>> exit_mm
>>>>> mmput => __mmput
>>>>> exit_mmap
>>>>> mmu_notifier_release
>>>>> mmu_notifier_ops->release
>>>>> kvm_arch_flush_shadow_all
>>>>> kvm_free_stage2_pgd
>>>>>
>>>>> In the first case, all references to the VM are dropped and hence none
>>>>> of the
>>>>> VCPU could still be executing.
>>>>>
>>>>> However, in the second case it may not be. So we have a potential
>>>>> problem with
>>>>> the VCPU trying to run even when the pages were unmapped. I think the
>>>>> root cause
>>>>> of all these issues boils down to the assumption that KVM holds a
>>>>> reference to
>>>>> MM (which is not necessarily the user space mapping. i.e mmgrab vs
>>>>> mmget).
>>>>> I am not sure if the VCPU should hold a reference to the mmaps to make
>>>>> sure
>>>>> it is safe to run. That way, the mmap stays until the VCPU eventually
>>>>> exits
>>>>> and we are safe all the way around.
>>>>
>>>>
>>>> Hmmm, my assumption is that if a VCPU is running, it means there is a
>>>> VCPU thread that shares the struct mm which is running, so I don't
>>>> understand how mmput would be able to call exit_mmap in the scenario
>>>> above?
>>>>
>>>> So the distinction here is that I don't assume that the VCPU fd holds a
>>>> reference to the mm, but I assume that the (running) VCPU thread does.
>>>> Is this incorrect?
>>>
>>>
>>> Sorry, I lost this thread in between.
>>
>>
>> No worries.
>>
>>>
>>> Hmm. You're right. The VCPU should have a refcount on mmap and it
>>> shouldn't
>>> do anything with the mmu if it has dropped it. I was confused based on an
>>> old bug report,[ See the description of commit 293f293637b55d "kvm-arm:
>>> Unmap shadow pagetables
>>> properly"], which was fixed.
>>>
>>
>
> This keeps haunting me :(. I happened to look at another report from Alex
> back in April,
> where a VCPU ends up in an mmu_notifier call back after the PGD was free'd !
> So it does
> look like you could end up in freeing the stage2 page table and yet another
> VCPU could
> be running.
>
> See : https://lkml.org/lkml/2017/4/21/820
>
> I am trying to see if I can reproduce the hypothesis.
>
I would also very much like to get to the bottom of this, and at the
very least try to get a valid explanation as to how a thread can be
*running* for a process where there are zero references to the struct
mm?

I guess I am asking where this mmput() can happen for a perfectly
running thread, which hasn't processes signals or exited itself yet.

The dump you reference above seems to indicate that it's happening
under memory pressure and trying to unmap memory from the VM to
allocate memory to the VM, but all seems to be happening within a VCPU
thread, or am I reading this wrong?

Thanks,
-Christoffer

\
 
 \ /
  Last update: 2017-07-17 16:46    [W:0.101 / U:2.844 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site