On 2/24/2016 1:07 PM, Ilya Maximets wrote:

On 23.02.2016 08:56, Xie, Huawei wrote:

quoted

On 2/22/2016 6:16 PM, Thomas Monjalon wrote:

quoted

2016-02-22 02:07, Xie, Huawei:

quoted

On 2/19/2016 5:05 PM, Ilya Maximets wrote:

quoted

On 19.02.2016 11:36, Xie, Huawei wrote:

quoted

On 2/19/2016 3:10 PM, Yuanhan Liu wrote:

quoted

On Fri, Feb 19, 2016 at 09:32:43AM +0300, Ilya Maximets wrote:

quoted

Signed-off-by: Ilya Maximets <redacted>
---
 doc/guides/prog_guide/thread_safety_dpdk_functions.rst | 1 +
 1 file changed, 1 insertion(+)

diff --git a/doc/guides/prog_guide/thread_safety_dpdk_functions.rst b/doc/guides/prog_guide/thread_safety_dpdk_functions.rst
index 403e5fc..13a6c89 100644
--- a/doc/guides/prog_guide/thread_safety_dpdk_functions.rst
+++ b/doc/guides/prog_guide/thread_safety_dpdk_functions.rst
 The mempool library is based on the DPDK lockless ring library and therefore is also multi-thread safe.
+rte_vhost_enqueue_burst() is also thread safe because based on lockless ring-buffer algorithm like the ring library.

FYI, Huawei meant to make rte_vhost_enqueue_burst() not be thread-safe,
to aligh with the usage of rte_eth_tx_burst().

	--yliu

I have a patch to remove the lockless enqueue. Unless there is strong
reason, i prefer vhost PMD to behave like other PMDs, with no internal
lockless algorithm. In future, for people who really need it, we could
have dynamic/static switch to enable it.

Thomas, what is your opinion on this and my patch removing lockless enqueue?

The thread safety behaviour is part of the API specification.
If we want to enable/disable such behaviour, it must be done with an API
function. But it would introduce a conditional statement in the fast path.
That's why the priority must be to keep a simple and consistent behaviour
and try to build around. An API complexity may be considered only if there
is a real (measured) gain.

Let us put the gain aside temporarily. I would do the measurement.
Vhost is wrapped as a PMD in Tetsuya's patch. And also in DPDK OVS's
case, it is wrapped as a vport like all other physical ports. The DPDK
app/OVS will treat all ports equally.

That is not true. Currently vhost in Open vSwitch implemented as a separate
netdev class. So, to use concurrency of vhost we just need to remove
2 lines (rte_spinlock_lock and rte_spinlock_unlock) from function
__netdev_dpdk_vhost_send(). This will not change behaviour of other types
of ports.

I checked OVS implementation. It raised several concerns.
For physical ports, it uses multiple queues to solve concurrent tx.
For vhost ports,  a) The thread safe behavior of vhost isn't used.
rte_spinlock is used outside. Yes, it could be removed.  b) If a packet
is send to vhost, it is directly enqueued to guest without buffering. We
could use thread safe ring to queue packets first and then enqueued to
guest at appropriate time later, then vhost internal lockless isn't needed.
Besides, IMO thread safe implementation adds the complexity of vhost
implementation.

quoted

It will add complexity if the app
needs to know that some supports concurrency while some not. Since all
other PMDs doesn't support thread safety, it doesn't make sense for
vhost PMD to support that. I believe the APP will not use that behavior.

quoted

From the API's point of view, if we previously implemented it wrongly,

we need to fix it as early as possible.