[ovs-dev] [PATCH] netdev-dpdk: Post-copy Live Migration support for vhost-user-client.

Maxime Coquelin maxime.coquelin at redhat.com
Tue May 14 11:36:48 UTC 2019



On 5/14/19 12:58 PM, Ilya Maximets wrote:
> On 14.05.2019 10:44, Maxime Coquelin wrote:
>> On 5/8/19 3:54 PM, Ilya Maximets wrote:
>>> From: Liliia Butorina <l.butorina at partner.samsung.com>
>>>
>>> Post-copy Live Migration for vHost supported since DPDK 18.11 and
>>> QEMU 2.12. New global config option 'vhost-postcopy-support' added
>>> to control this feature. Ex.:
>>>
>>>     ovs-vsctl set Open_vSwitch . other_config:vhost-postcopy-support=true
>>>
>>> Changing this value requires restarting the daemon. It's safe to
>>> enable this knob even if QEMU doesn't support post-copy LM.
>>>
>>> Feature marked as experimental and disabled by default because it may
>>> cause PMD thread hang on destination host on page fault for the time
>>> of page downloading from the source.
>>>
>>> Feature is not compatible with 'mlockall' and 'dequeue zero-copy'.
>>> Support added only for vhost-user-client.
>>>
>>> Signed-off-by: Liliia Butorina <l.butorina at partner.samsung.com>
>>> Co-authored-by: Ilya Maximets <i.maximets at samsung.com>
>>> Signed-off-by: Ilya Maximets <i.maximets at samsung.com>
>>
>> Thanks Ilya & Liliia for taking care of that.
>>
>>
>>> ---
>>>    Documentation/topics/dpdk/vhost-user.rst | 53 +++++++++++++++++++++++-
>>>    NEWS                                     |  1 +
>>>    lib/dpdk-stub.c                          |  6 +++
>>>    lib/dpdk.c                               | 13 ++++++
>>>    lib/dpdk.h                               |  1 +
>>>    lib/netdev-dpdk.c                        |  5 +++
>>>    vswitchd/vswitch.xml                     | 16 +++++++
>>>    7 files changed, 94 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/Documentation/topics/dpdk/vhost-user.rst b/Documentation/topics/dpdk/vhost-user.rst
>>> index 993797de5..6bec8c1f7 100644
>>> --- a/Documentation/topics/dpdk/vhost-user.rst
>>> +++ b/Documentation/topics/dpdk/vhost-user.rst
>>> @@ -111,7 +111,8 @@ the guest. There are two ways to do this: using QEMU directly, or using
>>>    libvirt.
>>>      .. note::
>>> -   IOMMU is not supported with vhost-user ports.
>>> +
>>> +   IOMMU and Post-copy Live Migration are not supported with vhost-user ports.
>>>      Adding vhost-user ports to the guest (QEMU)
>>>    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>> @@ -301,6 +302,52 @@ The default value is false.
>>>        QEMU). Starting with QEMU v2.9.1, vhost-iommu-support can safely be
>>>        enabled, even without having an IOMMU device, with no performance penalty.
>>>    +vhost-user-client Post-copy Live Migration Support (experimental)
>>> +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>> +
>>> +``Post-copy`` migration is the migration mode where the destination CPUs are
>>> +started before all the memory has been transferred. The main advantage is the
>>> +predictable migration time. Mostly used as a second phase after the normal
>>> +'pre-copy' migration in case it takes too long to converge.
>>> +
>>> +More information can be found in QEMU `docs`_.
>>> +
>>> +.. _`docs`: https://git.qemu.org/?p=qemu.git;a=blob;f=docs/devel/migration.rst
>>> +
>>> +Post-copy support may be enabled via a global config value
>>> +``vhost-postcopy-support``. Setting this to ``true`` enables Post-copy support
>>> +for all vhost-user-client ports::
>>> +
>>> +    $ ovs-vsctl set Open_vSwitch . other_config:vhost-postcopy-support=true
>>> +
>>> +The default value is ``false``.
>>> +
>>> +.. important::
>>> +
>>> +    Changing this value requires restarting the daemon.
>>> +
>>> +.. important::
>>> +
>>> +    DPDK Post-copy migration mode uses userfaultfd syscall to communicate with
>>> +    the kernel about page fault handling and uses shared memory based on huge
>>> +    pages. So destination host linux kernel should support userfaultfd over
>>> +    shared hugetlbfs. This feature only introduced in kernel upstream version
>>> +    4.11.
>>> +
>>> +    Post-copy feature supported in DPDK since 18.11.0 version and in QEMU
>>> +    since 2.12.0 version. But it's suggested to use QEMU >= 3.0.1 because
>>> +    migration recovery was fixed for post-copy in 3.0 and few additional bug
>>> +    fixes (like userfaulfd leak) was released in 3.0.1.
>>> +
>>> +    DPDK Post-copy feature requires avoiding to populate the guest memory
>>> +    (application must not call mlock* syscall). So enabling mlockall and
>>> +    dequeue zero-copy features is mis-compatible with post-copy feature.
>>> +
>>> +    Note that during migration of vhost-user device, PMD threads hang for the
>>> +    time of faulted pages download from source host. Transferring 1GB hugepage
>>> +    across a 10Gbps link possibly unacceptably slow. So recommended hugepage
>>> +    size is 2MB.
>>> +
>>>    .. _dpdk-testpmd:
>>>      DPDK in the Guest
>>> @@ -500,6 +547,10 @@ QEMU versions v2.10 and greater). This value can be set like so::
>>>      Because of this limitation, this feature is considered 'experimental'.
>>>    +.. note::
>>> +
>>> +   Post-copy Live Migration is not compatible with dequeue zero copy.
>>> +
>>>    Further information can be found in the
>>>    `DPDK documentation
>>>    <https://doc.dpdk.org/guides-18.11/prog_guide/vhost_lib.html>`__
>>> diff --git a/NEWS b/NEWS
>>> index 293531db0..f1f6f074e 100644
>>> --- a/NEWS
>>> +++ b/NEWS
>>> @@ -3,6 +3,7 @@ Post-v2.11.0
>>>       - DPDK:
>>>         * New option 'other_config:dpdk-socket-limit' to limit amount of
>>>           hugepage memory that can be used by DPDK.
>>> +     * Add support for vHost Post-copy Live Migration (experimental).
>>>       - OpenFlow:
>>>         * Removed support for OpenFlow 1.6 (draft), which ONF abandoned.
>>>         * New action "check_pkt_larger".
>>> diff --git a/lib/dpdk-stub.c b/lib/dpdk-stub.c
>>> index 1e0f46101..e55be5750 100644
>>> --- a/lib/dpdk-stub.c
>>> +++ b/lib/dpdk-stub.c
>>> @@ -56,6 +56,12 @@ dpdk_vhost_iommu_enabled(void)
>>>        return false;
>>>    }
>>>    +bool
>>> +dpdk_vhost_postcopy_enabled(void)
>>> +{
>>> +    return false;
>>> +}
>>> +
>>>    bool
>>>    dpdk_per_port_memory(void)
>>>    {
>>> diff --git a/lib/dpdk.c b/lib/dpdk.c
>>> index dc6171546..d9ec3cf64 100644
>>> --- a/lib/dpdk.c
>>> +++ b/lib/dpdk.c
>>> @@ -47,6 +47,8 @@ static FILE *log_stream = NULL;       /* Stream for DPDK log redirection */
>>>      static char *vhost_sock_dir = NULL;   /* Location of vhost-user sockets */
>>>    static bool vhost_iommu_enabled = false; /* Status of vHost IOMMU support */
>>> +static bool vhost_postcopy_enabled = false; /* Status of vHost POSTCOPY
>>> +                                             * support. */
>>>    static bool dpdk_initialized = false; /* Indicates successful initialization
>>>                                           * of DPDK. */
>>>    static bool per_port_memory = false; /* Status of per port memory support */
>>> @@ -316,6 +318,11 @@ dpdk_init__(const struct smap *ovs_other_config)
>>>        VLOG_INFO("Per port memory for DPDK devices %s.",
>>>                  per_port_memory ? "enabled" : "disabled");
>>>    +    vhost_postcopy_enabled = smap_get_bool(ovs_other_config,
>>> +                                           "vhost-postcopy-support", false);
>>> +    VLOG_INFO("POSTCOPY support for vhost-user-client %s.",
>>> +              vhost_postcopy_enabled ? "enabled" : "disabled");
>>> +
>>>        svec_add(&args, ovs_get_program_name());
>>>        construct_dpdk_args(ovs_other_config, &args);
>>>    @@ -492,6 +499,12 @@ dpdk_vhost_iommu_enabled(void)
>>>        return vhost_iommu_enabled;
>>>    }
>>>    +bool
>>> +dpdk_vhost_postcopy_enabled(void)
>>> +{
>>> +    return vhost_postcopy_enabled;
>>> +}
>>> +
>>>    bool
>>>    dpdk_per_port_memory(void)
>>>    {
>>> diff --git a/lib/dpdk.h b/lib/dpdk.h
>>> index bbb89d4e6..7dab83775 100644
>>> --- a/lib/dpdk.h
>>> +++ b/lib/dpdk.h
>>> @@ -39,6 +39,7 @@ void dpdk_init(const struct smap *ovs_other_config);
>>>    void dpdk_set_lcore_id(unsigned cpu);
>>>    const char *dpdk_get_vhost_sock_dir(void);
>>>    bool dpdk_vhost_iommu_enabled(void);
>>> +bool dpdk_vhost_postcopy_enabled(void);
>>>    bool dpdk_per_port_memory(void);
>>>    void print_dpdk_version(void);
>>>    void dpdk_status(const struct ovsrec_open_vswitch *);
>>> diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c
>>> index 47153dc60..c06f46931 100644
>>> --- a/lib/netdev-dpdk.c
>>> +++ b/lib/netdev-dpdk.c
>>> @@ -4147,6 +4147,11 @@ netdev_dpdk_vhost_client_reconfigure(struct netdev *netdev)
>>>                vhost_flags |= RTE_VHOST_USER_IOMMU_SUPPORT;
>>>            }
>>>    +        /* Enable POSTCOPY support, if explicitly requested. */
>>> +        if (dpdk_vhost_postcopy_enabled()) {
>>> +            vhost_flags |= RTE_VHOST_USER_POSTCOPY_SUPPORT;
>>> +        }
>>
>> Couldn't we also enforce postcopy disablement in the case --mlockall option is passed?
> 
> Sure. This could be done on 'dpdk_init' stage. I'll take care of this.
> 
> OTOH, this will block the case where source has mlock, but destination doesn't.

I would have thought source doesn't need to advertise
VHOST_USER_PROTOCOL_F_PAGEFAULT to have the postcopy migration to work.

I haven't tested it though, but in theory it shouldn't be needed.

> In theory, it's a valid case. Not sure how practical it is. What do you think?
Anyway, I don't think it would problematic in practice.

Thanks,
Maxime

> Best regards, Ilya Maximets.
> 


More information about the dev mailing list