Linux Virtualization - Oct 2018 - [virtio-dev] Re: [PATCH net-next v2 0/5] virtio: support packed ring

On Wed, Sep 12, 2018 at 12:16:32PM -0400, Michael S. Tsirkin
wrote:
> On Tue, Sep 11, 2018 at 01:37:26PM +0800, Tiwei Bie wrote:
> > On Mon, Sep 10, 2018 at 11:33:17AM +0800, Jason Wang wrote:
> > > On 2018?09?10? 11:00, Tiwei Bie wrote:
> > > > On Fri, Sep 07, 2018 at 09:00:49AM -0400, Michael S. Tsirkin
wrote:
> > > > > On Fri, Sep 07, 2018 at 09:22:25AM +0800, Tiwei Bie
wrote:
> > > > > > On Mon, Aug 27, 2018 at 05:00:40PM +0300, Michael
S. Tsirkin wrote:
> > > > > > > Are there still plans to test the performance
with vost pmd?
> > > > > > > vhost doesn't seem to show a performance
gain ...
> > > > > > > 
> > > > > > I tried some performance tests with vhost PMD. In
guest, the
> > > > > > XDP program will return XDP_DROP directly. And in
host, testpmd
> > > > > > will do txonly fwd.
> > > > > > 
> > > > > > When burst size is 1 and packet size is 64 in
testpmd and
> > > > > > testpmd needs to iterate 5 Tx queues (but only the
first two
> > > > > > queues are enabled) to prepare and inject packets,
I got ~12%
> > > > > > performance boost (5.7Mpps -> 6.4Mpps). And if
the vhost PMD
> > > > > > is faster (e.g. just need to iterate the first two
queues to
> > > > > > prepare and inject packets), then I got similar
performance
> > > > > > for both rings (~9.9Mpps) (packed ring's
performance can be
> > > > > > lower, because it's more complicated in
driver.)
> > > > > > 
> > > > > > I think packed ring makes vhost PMD faster, but it
doesn't make
> > > > > > the driver faster. In packed ring, the ring is
simplified, and
> > > > > > the handling of the ring in vhost (device) is also
simplified,
> > > > > > but things are not simplified in driver, e.g.
although there is
> > > > > > no desc table in the virtqueue anymore, driver
still needs to
> > > > > > maintain a private desc state table (which is
still managed as
> > > > > > a list in this patch set) to support the
out-of-order desc
> > > > > > processing in vhost (device).
> > > > > > 
> > > > > > I think this patch set is mainly to make the
driver have a full
> > > > > > functional support for the packed ring, which
makes it possible
> > > > > > to leverage the packed ring feature in vhost
(device). But I'm
> > > > > > not sure whether there is any other better idea,
I'd like to
> > > > > > hear your thoughts. Thanks!
> > > > > Just this: Jens seems to report a nice gain with virtio
and
> > > > > vhost pmd across the board. Try to compare virtio and
> > > > > virtio pmd to see what does pmd do better?
> > > > The virtio PMD (drivers/net/virtio) in DPDK doesn't need
to share
> > > > the virtio ring operation code with other drivers and is
highly
> > > > optimized for network. E.g. in Rx, the Rx burst function
won't
> > > > chain descs. So the ID management for the Rx ring can be
quite
> > > > simple and straightforward, we just need to initialize these
IDs
> > > > when initializing the ring and don't need to change
these IDs
> > > > in data path anymore (the mergable Rx code in that patch set
> > > > assumes the descs will be written back in order, which
should be
> > > > fixed. I.e., the ID in the desc should be used to index
vq->descx[]).
> > > > The Tx code in that patch set also assumes the descs will be
> > > > written back by device in order, which should be fixed.
> > > 
> > > Yes it is. I think I've pointed it out in some early version
of pmd patch.
> > > So I suspect part (or all) of the boost may come from in order
feature.
> > > 
> > > > 
> > > > But in kernel virtio driver, the virtio_ring.c is very
generic.
> > > > The enqueue (virtqueue_add()) and dequeue
(virtqueue_get_buf_ctx())
> > > > functions need to support all the virtio devices and should
be
> > > > able to handle all the possible cases that may happen. So
although
> > > > the packed ring can be very efficient in some cases,
currently
> > > > the room to optimize the performance in kernel's
virtio_ring.c
> > > > isn't that much. If we want to take the fully advantage
of the
> > > > packed ring's efficiency, we need some further e.g. API
changes
> > > > in virtio_ring.c, which shouldn't be part of this patch
set.
> > > 
> > > Could you please share more thoughts on this e.g how to improve
the API?
> > > Notice since the API is shared by both split ring and packed
ring, it may
> > > improve the performance of split ring as well. One can easily
imagine a
> > > batching API, but it does not have many real users now, the only
case is the
> > > XDP transmission which can accept an array of XDP frames.
> > 
> > I don't have detailed thoughts on this yet. But kernel's
> > virtio_ring.c is quite generic compared with what we did
> > in virtio PMD.
> 
> In what way? What are some things that aren't implemented there?
Below is the code corresponding to the virtqueue_add()
for Rx ring in virtio PMD:

https://github.com/DPDK/dpdk/blob/3605968c2fa7/drivers/net/virtio/virtio_rxtx.c#L278-L304

And below is the code of virtqueue_add() in Linux:

https://github.com/torvalds/linux/blob/54eda9df17f3/drivers/virtio/virtio_ring.c#L275-L417

In virtio PMD, for each packet (mbuf), the code is pretty
straightforward, it will just check whether there is one
available desc. If it's true, it will just fill this desc
directly.

But in virtqueue_add(), it's obvious that, the logic is
much more complicated or generic. It's supposed to be
able to handle sglist (which may consist of multiple IN
buffers and multiple OUT buffers at the same time), and
it will try to use indirect descriptors. Then it needs
several loops to parse the sglist. That's why I said
it's quite generic.
> 
> If what you say is true then we should take a careful look
> and not supporting these generic things with packed layout.
> Once we do support them it will be too late and we won't
> be able to get performance back.
I think it's a good point that we don't need to support
everything in packed ring (especially these which would
hurt the performance), as the packed ring aims at high
performance. I'm also wondering about the features. Is
there any possibility that we won't support the out of
order processing (at least not by default) in packed ring?
If I didn't miss anything, the need to support out of order
processing in packed ring will make the data structure
inside the driver not cache friendly which is similar to
the case of the descriptor table in the split ring (the
difference is that, it only happens in driver now).

> 
> 
> 
> > > 
> > > > So
> > > > I still think this patch set is mainly to make the kernel
virtio
> > > > driver to have a full functional support of the packed ring,
and
> > > > we can't expect impressive performance boost with it.
> > > 
> > > We can only gain when virtio ring layout is the bottleneck. If
there're
> > > bottlenecks elsewhere, we probably won't see any increasing
in the numbers.
> > > Vhost-net is an example, and lots of optimizations have proved
that virtio
> > > ring is not the main bottleneck for the current codes. I suspect
it also the
> > > case of virtio driver. Did perf tell us any interesting things in
virtio
> > > driver?
> > > 
> > > Thanks
> > > 
> > > > 
> > > > > 
> > > > > > > On Wed, Jul 11, 2018 at 10:27:06AM +0800,
Tiwei Bie wrote:
> > > > > > > > Hello everyone,
> > > > > > > > 
> > > > > > > > This patch set implements packed ring
support in virtio driver.
> > > > > > > > 
> > > > > > > > Some functional tests have been done
with Jason's
> > > > > > > > packed ring implementation in vhost:
> > > > > > > > 
> > > > > > > > https://lkml.org/lkml/2018/7/3/33
> > > > > > > > 
> > > > > > > > Both of ping and netperf worked as
expected.
> > > > > > > > 
> > > > > > > > v1 -> v2:
> > > > > > > > - Use READ_ONCE() to read event off_wrap
and flags together (Jason);
> > > > > > > > - Add comments related to ccw (Jason);
> > > > > > > > 
> > > > > > > > RFC (v6) -> v1:
> > > > > > > > - Avoid extra virtio_wmb() in
virtqueue_enable_cb_delayed_packed()
> > > > > > > >    when event idx is off (Jason);
> > > > > > > > - Fix bufs calculation in
virtqueue_enable_cb_delayed_packed() (Jason);
> > > > > > > > - Test the state of the desc at used_idx
instead of last_used_idx
> > > > > > > >    in
virtqueue_enable_cb_delayed_packed() (Jason);
> > > > > > > > - Save wrap counter (as part of queue
state) in the return value
> > > > > > > >    of
virtqueue_enable_cb_prepare_packed();
> > > > > > > > - Refine the packed ring definitions in
uapi;
> > > > > > > > - Rebase on the net-next tree;
> > > > > > > > 
> > > > > > > > RFC v5 -> RFC v6:
> > > > > > > > - Avoid tracking addr/len/flags when DMA
API isn't used (MST/Jason);
> > > > > > > > - Define wrap counter as bool (Jason);
> > > > > > > > - Use ALIGN() in vring_init_packed()
(Jason);
> > > > > > > > - Avoid using pointer to track `next` in
detach_buf_packed() (Jason);
> > > > > > > > - Add comments for barriers (Jason);
> > > > > > > > - Don't enable RING_PACKED on ccw
for now (noticed by Jason);
> > > > > > > > - Refine the memory barrier in
virtqueue_poll();
> > > > > > > > - Add a missing memory barrier in
virtqueue_enable_cb_delayed_packed();
> > > > > > > > - Remove the hacks in
virtqueue_enable_cb_prepare_packed();
> > > > > > > > 
> > > > > > > > RFC v4 -> RFC v5:
> > > > > > > > - Save DMA addr, etc in desc state
(Jason);
> > > > > > > > - Track used wrap counter;
> > > > > > > > 
> > > > > > > > RFC v3 -> RFC v4:
> > > > > > > > - Make ID allocation support
out-of-order (Jason);
> > > > > > > > - Various fixes for EVENT_IDX support;
> > > > > > > > 
> > > > > > > > RFC v2 -> RFC v3:
> > > > > > > > - Split into small patches (Jason);
> > > > > > > > - Add helper virtqueue_use_indirect()
(Jason);
> > > > > > > > - Just set id for the last descriptor of
a list (Jason);
> > > > > > > > - Calculate the prev in
virtqueue_add_packed() (Jason);
> > > > > > > > - Fix/improve desc suppression code
(Jason/MST);
> > > > > > > > - Refine the code layout for
XXX_split/packed and wrappers (MST);
> > > > > > > > - Fix the comments and API in uapi
(MST);
> > > > > > > > - Remove the BUG_ON() for indirect
(Jason);
> > > > > > > > - Some other refinements and bug fixes;
> > > > > > > > 
> > > > > > > > RFC v1 -> RFC v2:
> > > > > > > > - Add indirect descriptor support -
compile test only;
> > > > > > > > - Add event suppression supprt - compile
test only;
> > > > > > > > - Move vring_packed_init() out of uapi
(Jason, MST);
> > > > > > > > - Merge two loops into one in
virtqueue_add_packed() (Jason);
> > > > > > > > - Split vring_unmap_one() for packed
ring and split ring (Jason);
> > > > > > > > - Avoid using '%' operator
(Jason);
> > > > > > > > - Rename free_head -> next_avail_idx
(Jason);
> > > > > > > > - Add comments for virtio_wmb() in
virtqueue_add_packed() (Jason);
> > > > > > > > - Some other refinements and bug fixes;
> > > > > > > > 
> > > > > > > > Thanks!
> > > > > > > > 
> > > > > > > > Tiwei Bie (5):
> > > > > > > >    virtio: add packed ring definitions
> > > > > > > >    virtio_ring: support creating packed
ring
> > > > > > > >    virtio_ring: add packed ring support
> > > > > > > >    virtio_ring: add event idx support in
packed ring
> > > > > > > >    virtio_ring: enable packed ring
> > > > > > > > 
> > > > > > > >   drivers/s390/virtio/virtio_ccw.c   |  
14 +
> > > > > > > >   drivers/virtio/virtio_ring.c       |
1365 ++++++++++++++++++++++------
> > > > > > > >   include/linux/virtio_ring.h        |  
8 +-
> > > > > > > >   include/uapi/linux/virtio_config.h |  
3 +
> > > > > > > >   include/uapi/linux/virtio_ring.h   |  
43 +
> > > > > > > >   5 files changed, 1157 insertions(+),
276 deletions(-)
> > > > > > > > 
> > > > > > > > -- 
> > > > > > > > 2.18.0
> > > > > > >
---------------------------------------------------------------------
> > > > > > > To unsubscribe, e-mail:
virtio-dev-unsubscribe at lists.oasis-open.org
> > > > > > > For additional commands, e-mail:
virtio-dev-help at lists.oasis-open.org
> > > > > > > 
> > >

On 2018?09?13? 16:59, Tiwei Bie wrote:
>> If what you say is true then we should take a careful look
>> and not supporting these generic things with packed layout.
>> Once we do support them it will be too late and we won't
>> be able to get performance back.
> I think it's a good point that we don't need to support
> everything in packed ring (especially these which would
> hurt the performance), as the packed ring aims at high
> performance. I'm also wondering about the features. Is
> there any possibility that we won't support the out of
> order processing (at least not by default) in packed ring?
> If I didn't miss anything, the need to support out of order
> processing in packed ring will make the data structure
> inside the driver not cache friendly which is similar to
> the case of the descriptor table in the split ring (the
> difference is that, it only happens in driver now).
Out of order is not the only user, DMA is another one. We don't have 
used ring(len), so we need to maintain buffer length somewhere even for 
in order device. But if it's not too late, I second for a OUT_OF_ORDER 
feature. Starting from in order can have much simpler code in driver.

Thanks

On Thu, Sep 13, 2018 at 05:47:29PM +0800, Jason Wang
wrote:
> 
> 
> On 2018?09?13? 16:59, Tiwei Bie wrote:
> > > If what you say is true then we should take a careful look
> > > and not supporting these generic things with packed layout.
> > > Once we do support them it will be too late and we won't
> > > be able to get performance back.
> > I think it's a good point that we don't need to support
> > everything in packed ring (especially these which would
> > hurt the performance), as the packed ring aims at high
> > performance. I'm also wondering about the features. Is
> > there any possibility that we won't support the out of
> > order processing (at least not by default) in packed ring?
> > If I didn't miss anything, the need to support out of order
> > processing in packed ring will make the data structure
> > inside the driver not cache friendly which is similar to
> > the case of the descriptor table in the split ring (the
> > difference is that, it only happens in driver now).
> 
> Out of order is not the only user, DMA is another one. We don't have
used
> ring(len), so we need to maintain buffer length somewhere even for in order
> device.
For a bunch of systems dma unmap is a nop so we do not really
need to maintain it. It's a question of an API to detect that
and optimize for it. I posted a proposed patch for that -
want to try using that?
> But if it's not too late, I second for a OUT_OF_ORDER feature.
> Starting from in order can have much simpler code in driver.
> 
> Thanks
It's tricky to change the flag polarity because of compatibility
with legacy interfaces. Why is this such a big deal?

Let's teach drivers about IN_ORDER, then if devices
are in order it will get enabled by default.

-- 
MST