OVS DPDK DMA-Dev library/Design Discussion

[Bruce Richardson]

On Wed, Mar 30, 2022 at 11:25:05AM +0200, Maxime Coquelin wrote:
> 
> 
> On 3/30/22 04:02, Hu, Jiayu wrote:
> > 
> > 
> > > -----Original Message-----
> > > From: Ilya Maximets <i.maximets at ovn.org>
> > > Sent: Wednesday, March 30, 2022 1:45 AM
> > > To: Morten Brørup <mb at smartsharesystems.com>; Richardson, Bruce
> > > <bruce.richardson at intel.com>
> > > Cc: i.maximets at ovn.org; Maxime Coquelin <maxime.coquelin at redhat.com>;
> > > Van Haaren, Harry <harry.van.haaren at intel.com>; Pai G, Sunil
> > > <sunil.pai.g at intel.com>; Stokes, Ian <ian.stokes at intel.com>; Hu, Jiayu
> > > <jiayu.hu at intel.com>; Ferriter, Cian <cian.ferriter at intel.com>; ovs-
> > > dev at openvswitch.org; dev at dpdk.org; Mcnamara, John
> > > <john.mcnamara at intel.com>; O'Driscoll, Tim <tim.odriscoll at intel.com>;
> > > Finn, Emma <emma.finn at intel.com>
> > > Subject: Re: OVS DPDK DMA-Dev library/Design Discussion
> > > 
> > > On 3/29/22 19:13, Morten Brørup wrote:
> > > > > From: Bruce Richardson [mailto:bruce.richardson at intel.com]
> > > > > Sent: Tuesday, 29 March 2022 19.03
> > > > > 
> > > > > On Tue, Mar 29, 2022 at 06:45:19PM +0200, Morten Brørup wrote:
> > > > > > > From: Maxime Coquelin [mailto:maxime.coquelin at redhat.com]
> > > > > > > Sent: Tuesday, 29 March 2022 18.24
> > > > > > > 
> > > > > > > Hi Morten,
> > > > > > > 
> > > > > > > On 3/29/22 16:44, Morten Brørup wrote:
> > > > > > > > > From: Van Haaren, Harry [mailto:harry.van.haaren at intel.com]
> > > > > > > > > Sent: Tuesday, 29 March 2022 15.02
> > > > > > > > > 
> > > > > > > > > > From: Morten Brørup <mb at smartsharesystems.com>
> > > > > > > > > > Sent: Tuesday, March 29, 2022 1:51 PM
> > > > > > > > > > 
> > > > > > > > > > Having thought more about it, I think that a completely
> > > > > different
> > > > > > > architectural approach is required:
> > > > > > > > > > 
> > > > > > > > > > Many of the DPDK Ethernet PMDs implement a variety of RX and TX
> > > > > > > packet burst functions, each optimized for different CPU vector
> > > > > > > instruction sets. The availability of a DMA engine should be
> > > > > treated
> > > > > > > the same way. So I suggest that PMDs copying packet contents, e.g.
> > > > > > > memif, pcap, vmxnet3, should implement DMA optimized RX and TX
> > > > > packet
> > > > > > > burst functions.
> > > > > > > > > > 
> > > > > > > > > > Similarly for the DPDK vhost library.
> > > > > > > > > > 
> > > > > > > > > > In such an architecture, it would be the application's job to
> > > > > > > allocate DMA channels and assign them to the specific PMDs that
> > > > > should
> > > > > > > use them. But the actual use of the DMA channels would move down
> > > > > below
> > > > > > > the application and into the DPDK PMDs and libraries.
> > > > > > > > > > 
> > > > > > > > > > 
> > > > > > > > > > Med venlig hilsen / Kind regards, -Morten Brørup
> > > > > > > > > 
> > > > > > > > > Hi Morten,
> > > > > > > > > 
> > > > > > > > > That's *exactly* how this architecture is designed &
> > > > > implemented.
> > > > > > > > > 1.	The DMA configuration and initialization is up to the
> > > > > application
> > > > > > > (OVS).
> > > > > > > > > 2.	The VHost library is passed the DMA-dev ID, and its new
> > > > > async
> > > > > > > rx/tx APIs, and uses the DMA device to accelerate the copy.
> > > > > > > > > 
> > > > > > > > > Looking forward to talking on the call that just started.
> > > > > Regards, -
> > > > > > > Harry
> > > > > > > > > 
> > > > > > > > 
> > > > > > > > OK, thanks - as I said on the call, I haven't looked at the
> > > > > patches.
> > > > > > > > 
> > > > > > > > Then, I suppose that the TX completions can be handled in the TX
> > > > > > > function, and the RX completions can be handled in the RX function,
> > > > > > > just like the Ethdev PMDs handle packet descriptors:
> > > > > > > > 
> > > > > > > > TX_Burst(tx_packet_array):
> > > > > > > > 1.	Clean up descriptors processed by the NIC chip. --> Process
> > > > > TX
> > > > > > > DMA channel completions. (Effectively, the 2nd pipeline stage.)
> > > > > > > > 2.	Pass on the tx_packet_array to the NIC chip descriptors. --
> > > > > > Pass
> > > > > > > on the tx_packet_array to the TX DMA channel. (Effectively, the 1st
> > > > > > > pipeline stage.)
> > > > > > > 
> > > > > > > The problem is Tx function might not be called again, so enqueued
> > > > > > > packets in 2. may never be completed from a Virtio point of view.
> > > > > IOW,
> > > > > > > the packets will be copied to the Virtio descriptors buffers, but
> > > > > the
> > > > > > > descriptors will not be made available to the Virtio driver.
> > > > > > 
> > > > > > In that case, the application needs to call TX_Burst() periodically
> > > > > with an empty array, for completion purposes.
> > > > > > 
> > > > > > Or some sort of TX_Keepalive() function can be added to the DPDK
> > > > > library, to handle DMA completion. It might even handle multiple DMA
> > > > > channels, if convenient - and if possible without locking or other
> > > > > weird complexity.
> > > > > > 
> > > > > > Here is another idea, inspired by a presentation at one of the DPDK
> > > > > Userspace conferences. It may be wishful thinking, though:
> > > > > > 
> > > > > > Add an additional transaction to each DMA burst; a special
> > > > > transaction containing the memory write operation that makes the
> > > > > descriptors available to the Virtio driver.
> > > 
> > > I was talking with Maxime after the call today about the same idea.
> > > And it looks fairly doable, I would say.
> > 
> > If the idea is making DMA update used ring's index (2B) and packed ring descriptor's flag (2B),
> > yes, it will work functionally. But considering the offloading cost of DMA, it would hurt
> > performance. In addition, the latency of small copy of DMA is much higher than that of
> > CPU. So it will also increase latency.
> 
> I agree writing back descriptors using DMA can be sub-optimal,
> especially for packed ring where the head desc flags have to be written
> last.
> 

I think we'll have to try it out to check how it works. If we are already
doing hardware offload, adding one addition job to the DMA list may be a
very minor addition. 

[Incidentally, for something like a head-pointer update, using a fill
operation rather than a copy may be a good choice, as it avoid the need for
a memory read transaction from the DMA engine, since the data to be written
is already in the descriptor submitted.]

> Are you sure about latency? With current solution, the descriptors
> write-backs can happen quite some time after the DMA transfers are done,
> isn't it?
> 

For a polling receiver, having the DMA engine automatically do any
head-pointer updates after copies done should indeed lead to lowest
latency. For DMA engines that perform operations in parallel (such as Intel
DSA), we just need to ensure proper fencing of operations.

> > > 
> > > > > > 
> > > > > 
> > > > > That is something that can work, so long as the receiver is operating
> > > > > in polling mode. For cases where virtio interrupts are enabled, you
> > > > > still need to do a write to the eventfd in the kernel in vhost to
> > > > > signal the virtio side. That's not something that can be offloaded to
> > > > > a DMA engine, sadly, so we still need some form of completion call.
> > > > 
> > > > I guess that virtio interrupts is the most widely deployed scenario,
> > > > so let's ignore the DMA TX completion transaction for now - and call
> > > > it a possible future optimization for specific use cases. So it seems
> > > > that some form of completion call is unavoidable.
> > > > 
> > > 
> > > We could separate the actual kick of the guest with the data transfer.
> > > If interrupts are enabled, this means that the guest is not actively polling, i.e.
> > > we can allow some extra latency by performing the actual kick from the rx
> > > context, or, as Maxime said, if DMA engine can generate interrupts when the
> > > DMA queue is empty, vhost thread may listen to them and kick the guest if
> > > needed.  This will additionally remove the extra system call from the fast
> > > path.
> > 
> > Separating kick with data transfer is a very good idea. But it requires a dedicated
> > control plane thread to kick guest after DMA interrupt. Anyway, we can try this
> > optimization in the future.
> 
> Yes it requires a dedicated thread, but I don't think this is really an
> issue. Interrupt mode can be considered as slow-path.
> 

While not overly familiar with virtio/vhost, my main concern about
interrupt mode is not the handling of interrupts themselves when interrupt
mode is enabled, but rather ensuring that we have correct behaviour when
interrupt mode is enabled while copy operations are in-flight. Is it
possible for a guest to enable interrupt mode, receive packets but never be
woken up?

/Bruce