Re: [Qemu-devel] [RFC PATCH 00/20] Kemari for KVM v0.1

[Yoshiaki Tamura]

2010/4/22 Dor Laor <address@hidden>:
> On 04/22/2010 01:35 PM, Yoshiaki Tamura wrote:
>>
>> Dor Laor wrote:
>>>
>>> On 04/21/2010 08:57 AM, Yoshiaki Tamura wrote:
>>>>
>>>> Hi all,
>>>>
>>>> We have been implementing the prototype of Kemari for KVM, and we're
>>>> sending
>>>> this message to share what we have now and TODO lists. Hopefully, we
>>>> would like
>>>> to get early feedback to keep us in the right direction. Although
>>>> advanced
>>>> approaches in the TODO lists are fascinating, we would like to run
>>>> this project
>>>> step by step while absorbing comments from the community. The current
>>>> code is
>>>> based on qemu-kvm.git 2b644fd0e737407133c88054ba498e772ce01f27.
>>>>
>>>> For those who are new to Kemari for KVM, please take a look at the
>>>> following RFC which we posted last year.
>>>>
>>>> http://www.mail-archive.com/address@hidden/msg25022.html
>>>>
>>>> The transmission/transaction protocol, and most of the control logic is
>>>> implemented in QEMU. However, we needed a hack in KVM to prevent rip
>>>> from
>>>> proceeding before synchronizing VMs. It may also need some plumbing in
>>>> the
>>>> kernel side to guarantee replayability of certain events and
>>>> instructions,
>>>> integrate the RAS capabilities of newer x86 hardware with the HA
>>>> stack, as well
>>>> as for optimization purposes, for example.
>>>
>>> [ snap]
>>>
>>>>
>>>> The rest of this message describes TODO lists grouped by each topic.
>>>>
>>>> === event tapping ===
>>>>
>>>> Event tapping is the core component of Kemari, and it decides on which
>>>> event the
>>>> primary should synchronize with the secondary. The basic assumption
>>>> here is
>>>> that outgoing I/O operations are idempotent, which is usually true for
>>>> disk I/O
>>>> and reliable network protocols such as TCP.
>>>
>>> IMO any type of network even should be stalled too. What if the VM runs
>>> non tcp protocol and the packet that the master node sent reached some
>>> remote client and before the sync to the slave the master failed?
>>
>> In current implementation, it is actually stalling any type of network
>> that goes through virtio-net.
>>
>> However, if the application was using unreliable protocols, it should
>> have its own recovering mechanism, or it should be completely stateless.
>
> Why do you treat tcp differently? You can damage the entire VM this way -
> think of dhcp request that was dropped on the moment you switched between
> the master and the slave?

I'm not trying to say that we should treat tcp differently, but just
it's severe.
In case of dhcp request, the client would have a chance to retry after
failover, correct?
BTW, in current implementation, it's synchronizing before dhcp ack is sent.
But in case of tcp, once you send ack to the client before sync, there
is no way to recover.

>>> [snap]
>>>
>>>
>>>> === clock ===
>>>>
>>>> Since synchronizing the virtual machines every time the TSC is
>>>> accessed would be
>>>> prohibitive, the transmission of the TSC will be done lazily, which
>>>> means
>>>> delaying it until there is a non-TSC synchronization point arrives.
>>>
>>> Why do you specifically care about the tsc sync? When you sync all the
>>> IO model on snapshot it also synchronizes the tsc.
>
> So, do you agree that an extra clock synchronization is not needed since it
> is done anyway as part of the live migration state sync?

I agree that its sent as part of the live migration.
What I wanted to say here is that this is not something for real time
applications.
I usually get questions like can this guarantee fault tolerance for
real time applications.

>>> In general, can you please explain the 'algorithm' for continuous
>>> snapshots (is that what you like to do?):
>>
>> Yes, of course.
>> Sorry for being less informative.
>>
>>> A trivial one would we to :
>>> - do X online snapshots/sec
>>
>> I currently don't have good numbers that I can share right now.
>> Snapshots/sec depends on what kind of workload is running, and if the
>> guest was almost idle, there will be no snapshots in 5sec. On the other
>> hand, if the guest was running I/O intensive workloads (netperf, iozone
>> for example), there will be about 50 snapshots/sec.
>>
>>> - Stall all IO (disk/block) from the guest to the outside world
>>> until the previous snapshot reaches the slave.
>>
>> Yes, it does.
>>
>>> - Snapshots are made of
>>
>> Full device model + diff of dirty pages from the last snapshot.
>>
>>> - diff of dirty pages from last snapshot
>>
>> This also depends on the workload.
>> In case of I/O intensive workloads, dirty pages are usually less than 100.
>
> The hardest would be memory intensive loads.
> So 100 snap/sec means latency of 10msec right?
> (not that it's not ok, with faster hw and IB you'll be able to get much
> more)

Doesn't 100 snap/sec mean the interval of snap is 10msec?
IIUC, to get the latency, you need to get, Time to transfer VM + Time
to get response from the receiver.

It's hard to say which load is the hardest.
Memory intensive load, who don't generate I/O often, will suffer from
long sync time for that moment, but would have chances to continue its
process until sync.
I/O intensive load, who don't dirty much pages, will suffer from
getting VPU stopped often, but its sync time is relatively shorter.

>>> - Qemu device model (+kvm's) diff from last.
>>
>> We're currently sending full copy because we're completely reusing this
>> part of existing live migration framework.
>>
>> Last time we measured, it was about 13KB.
>> But it varies by which QEMU version is used.
>>
>>> You can do 'light' snapshots in between to send dirty pages to reduce
>>> snapshot time.
>>
>> I agree. That's one of the advanced topic we would like to try too.
>>
>>> I wrote the above to serve a reference for your comments so it will map
>>> into my mind. Thanks, dor
>>
>> Thank your for the guidance.
>> I hope this answers to your question.
>>
>> At the same time, I would also be happy it we could discuss how to
>> implement too. In fact, we needed a hack to prevent rip from proceeding
>> in KVM, which turned out that it was not the best workaround.
>
> There are brute force solutions like
> - stop the guest until you send all of the snapshot to the remote (like
>  standard live migration)

We've implemented this way so far.

> - Stop + fork + cont the father
>
> Or mark the recent dirty pages that were not sent to the remote as write
> protected and copy them if touched.

I think I had that suggestion from Avi before.
And yes, it's very fascinating.

Meanwhile, if you look at the diffstat, it needed to touch many parts of QEMU.
Before going into further implementation, I wanted to check that I'm
in the right track for doing this project.


>> Thanks,
>>
>> Yoshi
>>
>>>
>>>>
>>>> TODO:
>>>> - Synchronization of clock sources (need to intercept TSC reads, etc).
>>>>
>>>> === usability ===
>>>>
>>>> These are items that defines how users interact with Kemari.
>>>>
>>>> TODO:
>>>> - Kemarid daemon that takes care of the cluster management/monitoring
>>>> side of things.
>>>> - Some device emulators might need minor modifications to work well
>>>> with Kemari. Use white(black)-listing to take the burden of
>>>> choosing the right device model off the users.
>>>>
>>>> === optimizations ===
>>>>
>>>> Although the big picture can be realized by completing the TODO list
>>>> above, we
>>>> need some optimizations/enhancements to make Kemari useful in real
>>>> world, and
>>>> these are items what needs to be done for that.
>>>>
>>>> TODO:
>>>> - SMP (for the sake of performance might need to implement a
>>>> synchronization protocol that can maintain two or more
>>>> synchronization points active at any given moment)
>>>> - VGA (leverage VNC's subtilting mechanism to identify fb pages that
>>>> are really dirty).
>>>>
>>>>
>>>> Any comments/suggestions would be greatly appreciated.
>>>>
>>>> Thanks,
>>>>
>>>> Yoshi
>>>>
>>>> --
>>>>
>>>> Kemari starts synchronizing VMs when QEMU handles I/O requests.
>>>> Without this patch VCPU state is already proceeded before
>>>> synchronization, and after failover to the VM on the receiver, it
>>>> hangs because of this.
>>>>
>>>> Signed-off-by: Yoshiaki Tamura<address@hidden>
>>>> ---
>>>> arch/x86/include/asm/kvm_host.h | 1 +
>>>> arch/x86/kvm/svm.c | 11 ++++++++---
>>>> arch/x86/kvm/vmx.c | 11 ++++++++---
>>>> arch/x86/kvm/x86.c | 4 ++++
>>>> 4 files changed, 21 insertions(+), 6 deletions(-)
>>>>
>>>> diff --git a/arch/x86/include/asm/kvm_host.h
>>>> b/arch/x86/include/asm/kvm_host.h
>>>> index 26c629a..7b8f514 100644
>>>> --- a/arch/x86/include/asm/kvm_host.h
>>>> +++ b/arch/x86/include/asm/kvm_host.h
>>>> @@ -227,6 +227,7 @@ struct kvm_pio_request {
>>>> int in;
>>>> int port;
>>>> int size;
>>>> + bool lazy_skip;
>>>> };
>>>>
>>>> /*
>>>> diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
>>>> index d04c7ad..e373245 100644
>>>> --- a/arch/x86/kvm/svm.c
>>>> +++ b/arch/x86/kvm/svm.c
>>>> @@ -1495,7 +1495,7 @@ static int io_interception(struct vcpu_svm *svm)
>>>> {
>>>> struct kvm_vcpu *vcpu =&svm->vcpu;
>>>> u32 io_info = svm->vmcb->control.exit_info_1; /* address size bug? */
>>>> - int size, in, string;
>>>> + int size, in, string, ret;
>>>> unsigned port;
>>>>
>>>> ++svm->vcpu.stat.io_exits;
>>>> @@ -1507,9 +1507,14 @@ static int io_interception(struct vcpu_svm *svm)
>>>> port = io_info>> 16;
>>>> size = (io_info& SVM_IOIO_SIZE_MASK)>> SVM_IOIO_SIZE_SHIFT;
>>>> svm->next_rip = svm->vmcb->control.exit_info_2;
>>>> - skip_emulated_instruction(&svm->vcpu);
>>>>
>>>> - return kvm_fast_pio_out(vcpu, size, port);
>>>> + ret = kvm_fast_pio_out(vcpu, size, port);
>>>> + if (ret)
>>>> + skip_emulated_instruction(&svm->vcpu);
>>>> + else
>>>> + vcpu->arch.pio.lazy_skip = true;
>>>> +
>>>> + return ret;
>>>> }
>>>>
>>>> static int nmi_interception(struct vcpu_svm *svm)
>>>> diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
>>>> index 41e63bb..09052d6 100644
>>>> --- a/arch/x86/kvm/vmx.c
>>>> +++ b/arch/x86/kvm/vmx.c
>>>> @@ -2975,7 +2975,7 @@ static int handle_triple_fault(struct kvm_vcpu
>>>> *vcpu)
>>>> static int handle_io(struct kvm_vcpu *vcpu)
>>>> {
>>>> unsigned long exit_qualification;
>>>> - int size, in, string;
>>>> + int size, in, string, ret;
>>>> unsigned port;
>>>>
>>>> exit_qualification = vmcs_readl(EXIT_QUALIFICATION);
>>>> @@ -2989,9 +2989,14 @@ static int handle_io(struct kvm_vcpu *vcpu)
>>>>
>>>> port = exit_qualification>> 16;
>>>> size = (exit_qualification& 7) + 1;
>>>> - skip_emulated_instruction(vcpu);
>>>>
>>>> - return kvm_fast_pio_out(vcpu, size, port);
>>>> + ret = kvm_fast_pio_out(vcpu, size, port);
>>>> + if (ret)
>>>> + skip_emulated_instruction(vcpu);
>>>> + else
>>>> + vcpu->arch.pio.lazy_skip = true;
>>>> +
>>>> + return ret;
>>>> }
>>>>
>>>> static void
>>>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>>>> index fd5c3d3..cc308d2 100644
>>>> --- a/arch/x86/kvm/x86.c
>>>> +++ b/arch/x86/kvm/x86.c
>>>> @@ -4544,6 +4544,10 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu
>>>> *vcpu, struct kvm_run *kvm_run)
>>>> if (!irqchip_in_kernel(vcpu->kvm))
>>>> kvm_set_cr8(vcpu, kvm_run->cr8);
>>>>
>>>> + if (vcpu->arch.pio.lazy_skip)
>>>> + kvm_x86_ops->skip_emulated_instruction(vcpu);
>>>> + vcpu->arch.pio.lazy_skip = false;
>>>> +
>>>> if (vcpu->arch.pio.count || vcpu->mmio_needed ||
>>>> vcpu->arch.emulate_ctxt.restart) {
>>>> if (vcpu->mmio_needed) {
>>>
>>>
>>>
>>>
>>
>>
>>
>
>
>
>