Re: [PATCH] vhost: make vhost_zerocopy_callback more efficient by poll_queue base on vhost status

[Jason Wang <jasowang@redhat.com>]

On 02/25/2014 04:56 PM, Qin Chuanyu wrote:
> On 2014/2/25 16:13, Jason Wang wrote:
>> On 02/25/2014 03:53 PM, Qin Chuanyu wrote:
>>> On 2014/2/25 15:38, Jason Wang wrote:
>>>> On 02/25/2014 02:55 PM, Qin Chuanyu wrote:
>>>>> guest kick vhost base on vring flag status and get perfermance
>>>>> improved,
>>>>> vhost_zerocopy_callback could do this in the same way, as
>>>>> virtqueue_enable_cb need one more check after change the status of
>>>>> avail_ring flags, vhost also do the same thing after
>>>>> vhost_enable_notify
>>>>>
>>>>> test result list as below:
>>>>> guest and host: suse11sp3, netperf, intel 2.4GHz
>>>>> +-------+----------+---------+----------+---------+
>>>>> |       |   old              |   new              |
>>>>> +-------+----------+---------+----------+---------+
>>>>> |  UDP  |  Gbit/s  |   PPS   |  Gbit/s  |   PPS   |
>>>>> |  256  | 0.74805  | 321309  | 0.77933  | 334743  |
>>>>> |  512  |   1.42   | 328475  |   1.44   | 333550  |
>>>>> | 1024  |   2.79   | 334426  |   2.81   | 336986  |
>>>>> | 1460  |   3.71   | 316215  |   4.02   | 342325  |
>>>>> +-------+----------+---------+----------+---------+
>>>>
>>>> Looks good, do you have cpu utilization number?
>>> +------+----------+--------+----------+--------+--------+---------+
>>> |      |             old              |            new            |
>>> +------+----------+--------+----------+--------+--------+---------+
>>> | UDP  |  Gbit/s  |  PPS   |CPU idle% | Gbit/s |   PPS  |CPU idle%|
>>> | 256  | 0.74805  | 321309 |  87.16   | 0.77933| 334743 |  90.71  |
>>> | 512  |   1.42   | 328475 |  87.03   |  1.44  | 333550 |  90.43  |
>>> | 1024 |   2.79   | 334426 |  89.09   |  2.81  | 336986 |  89.55  |
>>> | 1460 |   3.71   | 316215 |  87.53   |  4.02  | 342325 |  89.58  |
>>> +------+----------+--------+----------+--------+--------+---------+
>>> after change, less cpu has been used.
>>
>> Thanks
>>>>>
>>>>> Signed-off-by: Chuanyu Qin <qinchuanyu@huawei.com>
>>>>> ---
>>>>>    drivers/vhost/net.c |   10 +++++++++-
>>>>>    1 files changed, 9 insertions(+), 1 deletions(-)
>>>>>
>>>>> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
>>>>> index a0fa5de..9bc0a15 100644
>>>>> --- a/drivers/vhost/net.c
>>>>> +++ b/drivers/vhost/net.c
>>>>> @@ -322,7 +322,9 @@ static void vhost_zerocopy_callback(struct
>>>>> ubuf_info *ubuf, bool success)
>>>>>         * (the value 16 here is more or less arbitrary, it's tuned to
>>>>> trigger
>>>>>         * less than 10% of times).
>>>>>         */
>>>>> -    if (cnt <= 1 || !(cnt % 16))
>>>>> +    smp_rmb();
>>>>
>>>> Better add a comment to explain why this is needed.
>>>>
>>>> Looks like what you need is a smp_mb() here to make sure the len is
>>>> updated before testing vq->used_flags?
>>> I wanner make sure the used_flags is updated, is smp_rmb() enough?
>>> or a smp_mb() is needed?
>>
>> used_flags was guaranteed to be updated after smp_mb() in
>> vhost_enable_notify(). And vhost_net_ubuf_put() does a
>> atomic_sub_return() which implements memory barrier. So looks like
>> there's no need for an extra one.
>>
> atomic_sub_return only do as below:
>     asm volatile(LOCK_PREFIX "xaddl %0, %1"
>              : "+r" (i), "+m" (v->counter)
>              : : "memory");
>     return i + __i;
>
> google as below:
> The "memory" clobber makes GCC assume that any memory may be
> arbitrarily read or written by the asm block, so will prevent the
> compiler from reordering loads or stores across it:
>
>     This will cause GCC to not keep memory values cached in registers
> across the assembler instruction and not optimize stores or loads to
> that memory.
>
> (That does not prevent a CPU from reordering loads and stores with
> respect to another CPU, though; you need real memory barrier
> instructions for that.)

Ok, I thought lock prefix should do this but recheck the manual, it was
not a serialize instruction. 
> -----------------------------------------------------------------
> vhost_zerocopy_callback might be involved in softirq context in another
> cpu ,so I think smp_rmb() is still needed, is it right ?
>

rmb is only used to prevent cpu from reordering read. You need a mb here
to prevent the used_flags to be checked before updating the len.
Otherwise the you may lost a vhost_poll() consider:
                                   
vhost_disable_notify()
[CPU0]                                                             
if ((!(vq->used_flags & VRING_USED_F_NO_NOTIFY)) [CPU1]
vhost_enable_notify() [CPU0]
vhost_zerocopy_singal_used() [CPU0]
vq->heads[ubuf->desc].len = success ? [CPU1]


So I think you need a smp_wmb() here.
>>>>> +    if ((!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
>>>>> +            && (cnt <= 1 || !(cnt % 16)))
>>>>>            vhost_poll_queue(&vq->poll);
>>>>>
>>>>>        rcu_read_unlock_bh();
>>>>> @@ -386,6 +388,12 @@ static void handle_tx(struct vhost_net *net)
>>>>>                    vhost_disable_notify(&net->dev, vq);
>>>>>                    continue;
>>>>>                }
>>>>> +            /* there might skb been freed between last
>>>>> +            * vhost_zerocopy_signal_used and vhost_enable_notify,
>>>>> +            * so one more check is needed.
>>>>> +            */
>>>>> +            if (zcopy)
>>>>> +                vhost_zerocopy_signal_used(net, vq);
>>>>>                break;
>>>>>            }
>>>>>            if (in) {
>>>>
>>>>
>>>> .
>>>>
>>>
>>>
>>> -- 
>>> To unsubscribe from this list: send the line "unsubscribe kvm" in
>>> the body of a message to majordomo@vger.kernel.org
>>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>>
>>
>> .
>>
>
>
> -- 
> To unsubscribe from this list: send the line "unsubscribe kvm" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html