[dpdk-dev] data copy in vhost-user
Zoltan, Huawei, Thank you for your replies. -- Best regards, Nikita Kalyazin, n.kalyazin at samsung.com CE OS Group Samsung R Institute Russia Tel: +7 (495) 797-25-00 #3816 Tel: +7 (495) 797-25-03 Office #1501, 12-1, Dvintsev str., Moscow, 127018, Russia On Tue, Apr 28, 2015 at 01:24:25PM +0100, Zoltan Kiss wrote: > > > On 28/04/15 02:22, Xie, Huawei wrote: > > > > > >> -Original Message- > >> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Zoltan Kiss > >> Sent: Tuesday, April 28, 2015 12:27 AM > >> To: Nikita Kalyazin; dev at dpdk.org > >> Subject: Re: [dpdk-dev] data copy in vhost-user > >> > >> > >> > >> On 27/04/15 12:54, Nikita Kalyazin wrote: > >>> Hi, > >>> > >>> > >>> As far as I understand, DPDK vhost-user implementation requires data copy > >> for either RX or TX (rte_vhost_dequeue_burst() and > >> rte_vhost_enqueue_burst()). It means that two data copies are needed to > >> transfer a packet from one VM to another. > >>> > >>> Why is not it possible to eliminate one of the copies (e.g., > >> rte_vhost_enqueue_burst() might set up a reference at vring descriptor to > >> mbuf's data rather than copying the data)? > > This had been added to the to-do list. We could delay the copy until the > > real copy is needed. > >> > >> I'm just guessing, but in case of VM-to-VM traffic the receiving one > >> could hold onto the buffer indefinitely, preventing the sender to reuse > >> the buffer. That could lead to a DoS in some cases, and shutting down > >> the sender would be also tricky. At least in case of Xen > >> netback/netfront that's the reason. A reasonable solution for this > >> problem is to make sure the buffer is swapped out with a copy after a > >> finite time. > > Do you mean we associate a timeout for the buffer? > Yes, I think xen-netback had such a version once, but it was removed. As > far as I know the overhead and complexity of handling these timeouts > were too severe. > I might be wrong about this, I don't know if this problem applies here > as well or not. > > >> > >> Regards, > >> > >> Zoltan
[dpdk-dev] data copy in vhost-user
On 28/04/15 02:22, Xie, Huawei wrote: > > >> -Original Message- >> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Zoltan Kiss >> Sent: Tuesday, April 28, 2015 12:27 AM >> To: Nikita Kalyazin; dev at dpdk.org >> Subject: Re: [dpdk-dev] data copy in vhost-user >> >> >> >> On 27/04/15 12:54, Nikita Kalyazin wrote: >>> Hi, >>> >>> >>> As far as I understand, DPDK vhost-user implementation requires data copy >> for either RX or TX (rte_vhost_dequeue_burst() and >> rte_vhost_enqueue_burst()). It means that two data copies are needed to >> transfer a packet from one VM to another. >>> >>> Why is not it possible to eliminate one of the copies (e.g., >> rte_vhost_enqueue_burst() might set up a reference at vring descriptor to >> mbuf's data rather than copying the data)? > This had been added to the to-do list. We could delay the copy until the real > copy is needed. >> >> I'm just guessing, but in case of VM-to-VM traffic the receiving one >> could hold onto the buffer indefinitely, preventing the sender to reuse >> the buffer. That could lead to a DoS in some cases, and shutting down >> the sender would be also tricky. At least in case of Xen >> netback/netfront that's the reason. A reasonable solution for this >> problem is to make sure the buffer is swapped out with a copy after a >> finite time. > Do you mean we associate a timeout for the buffer? Yes, I think xen-netback had such a version once, but it was removed. As far as I know the overhead and complexity of handling these timeouts were too severe. I might be wrong about this, I don't know if this problem applies here as well or not. >> >> Regards, >> >> Zoltan
[dpdk-dev] data copy in vhost-user
> -Original Message- > From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Zoltan Kiss > Sent: Tuesday, April 28, 2015 12:27 AM > To: Nikita Kalyazin; dev at dpdk.org > Subject: Re: [dpdk-dev] data copy in vhost-user > > > > On 27/04/15 12:54, Nikita Kalyazin wrote: > > Hi, > > > > > > As far as I understand, DPDK vhost-user implementation requires data copy > for either RX or TX (rte_vhost_dequeue_burst() and > rte_vhost_enqueue_burst()). It means that two data copies are needed to > transfer a packet from one VM to another. > > > > Why is not it possible to eliminate one of the copies (e.g., > rte_vhost_enqueue_burst() might set up a reference at vring descriptor to > mbuf's data rather than copying the data)? This had been added to the to-do list. We could delay the copy until the real copy is needed. > > I'm just guessing, but in case of VM-to-VM traffic the receiving one > could hold onto the buffer indefinitely, preventing the sender to reuse > the buffer. That could lead to a DoS in some cases, and shutting down > the sender would be also tricky. At least in case of Xen > netback/netfront that's the reason. A reasonable solution for this > problem is to make sure the buffer is swapped out with a copy after a > finite time. Do you mean we associate a timeout for the buffer? > > Regards, > > Zoltan
[dpdk-dev] data copy in vhost-user
On 27/04/15 12:54, Nikita Kalyazin wrote: > Hi, > > > As far as I understand, DPDK vhost-user implementation requires data copy for > either RX or TX (rte_vhost_dequeue_burst() and rte_vhost_enqueue_burst()). It > means that two data copies are needed to transfer a packet from one VM to > another. > > Why is not it possible to eliminate one of the copies (e.g., > rte_vhost_enqueue_burst() might set up a reference at vring descriptor to > mbuf's data rather than copying the data)? I'm just guessing, but in case of VM-to-VM traffic the receiving one could hold onto the buffer indefinitely, preventing the sender to reuse the buffer. That could lead to a DoS in some cases, and shutting down the sender would be also tricky. At least in case of Xen netback/netfront that's the reason. A reasonable solution for this problem is to make sure the buffer is swapped out with a copy after a finite time. Regards, Zoltan
