On Wed, Mar 25, 2020 at 14:36 AM Karlsson, Magnus
magnus.karlsson@xxxxxxxxx> wrote:
> On Wed, Mar 25, 2020 at 1:40 PM Gaul, Maximilian <maximilian.gaul@xxxxxx> wrote:
> >
> > On Wed, Mar 25, 2020 at 12:04 AM Karlsson, Magnus
> > magnus.karlsson@xxxxxxxxx> wrote:
> >
> > > On Wed, Mar 25, 2020 at 11:45 AM Gaul, Maximilian
> > > <maximilian.gaul@xxxxxx> wrote:
> > > >
> > > > On Wed, Mar 25, 2020 at 11:24 AM Karlsson, Magnus
> > > > <magnus.karlsson@xxxxxxxxx> wrote:
> > > >
> > > > > On Wed, Mar 25, 2020 at 11:02 AM Gaul, Maximilian
> > > > > <maximilian.gaul@xxxxxx> wrote:
> > > > > >
> > > > > > On Wed, Mar 25, 2020 at 10:41 AM Karlsson, Magnus
> > > > > > <magnus.karlsson@xxxxxxxxx> wrote:
> > > > > >
> > > > > > > On Wed, Mar 25, 2020 at 10:04 AM Gaul, Maximilian
> > > > > > > <maximilian.gaul@xxxxxx> wrote:
> > > > > > > >
> > > > > > > > I am running a Multi-AF-XDP-Socket approach per RX-Queue (using Shared Umem).
> > > > > > > >
> > > > > > > > Unfortunately I am noticing, that at around 650k pps, the *ksoftirqd*-thread of that RX-Queue ramps up to 100% thus leading to packet loss.
> > > > > > > > I tried setting *XDP_USE_NEED_WAKEUP* on *xsk_socket_cfg.bind_flags* but those bind_flags are only taken into account if *umem->refcount > 1* (libbpf/xsk.c - xsk_socket__create()).
> > > > > > > > As far as I understand this correctly, only the first socket is able to set *XDP_USE_NEED_WAKEUP* because for all sockets after, *umem->refcount* is going to be at least 2.
> > > > > > >
> > > > > > > Yes, the other sockets just inherit the settings of the first one.
> > > > > > >
> > > > > > > Are you using the SKB mode? What is your packet size? Sounds like a
> > > > > > > low number unless you have large packets and are using the SKB mode.
> > > > > > >
> > > > > >
> > > > > > These are the flags I set right before calling `xsk_socket__create`:
> > > > > >
> > > > > > xsk_socket_cfg.xdp_flags = cfg->xdp_flags | XDP_FLAGS_DRV_MODE | XDP_ZEROCOPY;
> > > > > > xsk_socket_cfg.bind_flags = cfg->xsk_bind_flags | XDP_USE_NEED_WAKEUP;
> > > > >
> > > > > XDP_ZEROCOPY is a bind flag not an XDP flag, so please move it there.
> > > > > If you get an error when you have it set, it means that your setup
> > > > > does not support zero-copy for some reason. Check what kernel version
> > > > > you are using and the the driver you are using supports zero-copy. I
> > > > > believe you need to use a queue id>=32 in the Mellanox driver for it
> > > > > to work in zero-copy mode. Below 32, you will get copy mode.
> > > > >
> > > > > > Packet size is around 1492 bytes.
> > > > >
> > > > > Seems that you are using SKB mode then, not zero-copy.
> > > > >
> > > > > /Magnus
> > > > >
> > > >
> > > > Thank you for the hint. As you correctly said, I get an error if I use *XDP_ZEROCOPY*. But as far as I understand, packet rates should be higher in Driver-Mode even without zero-copy?
> > >
> > > Yes, I would expect that too.
> > >
> > > > I just updated to the latest driver- and firmware version:
> > > >
> > > > $ sudo ethtool -i <if>
> > > > driver: mlx5_core
> > > > version: 5.0-0
> > > > firmware-version: 16.27.1016 (MT_0000000012)
> > >
> > > What kernel version are you using? And you should use the driver from
> > > that same kernel.
> > >
> >
> > I am using
> >
> > $ uname -a
> > Linux 5.4.0-4-amd64 #1 SMP Debian 5.4.19-1 (2020-02-13) x86_64 GNU/Linux
> >
> > At the moment, Mellanox only supports Debian until version 10.0 (https://www.mellanox.com/products/infiniband-drivers/linux/mlnx_ofed) which is Kernel 4.19.
> > But because in Kernel 4.19, not all AF-XDP features are available, I had to upgrade. I was not sure which Kernel-Version would be the minimum in order to be able to use AF-XDP completely so I went with 5.4.
> > Installation was successfull (with *--skip-distro-check*) so I thought this should work?
>
> You better contact somebody from Mellanox for this info. I do not
> know. But Mellanox has zero-copy support in kernel 5.4.
>
> > > > I actually have to correct myself: Incomming packets are 1442 bytes.
> > > > Can you give me the link between packet size and whether the NIC is running in SKB or DRV mode?
> > >
> > > Sorry, do not understand this. Could you please elaborate?
> > >
> >
> > You answered to my reply that packets are 1492 bytes "Seems that you are using SKB mode then, not zero-copy." so because of this I thought there is a relation between packet size and SKB mode?
>
> There is no relationship between SKB mode and packet size. They are
> orthogonal. Though there is a relationship between packet size and
> performance and of course SKB mode vs zero-copy mode and performance.
>
> > > > Mr. Brouer held a talk about (https://people.netfilter.org/hawk/presentations/driving-IT2017/driving-IT-2017_XDP_eBPF_technology_Jesper_Brouer.pdf) about XDP, mentioning in slide 11/27 that *mlx5 (v4.9)* has native XDP support.
> > >
> > > Yes, but only if you use queue id>= 32. What queue id are you binding to?
> > >
> >
> > Usually it is queue 0 but I also tried queue 32 and queue 36 - but that didn't change anything on the behavior.
>
> It should make a difference, if I remember correctly, but somebody
> from Mellanox certainly knows better. Try sending a mail to Maxim who
> wrote the Mellanox driver support. Maxim Mikityanskiy
> <maximmi@xxxxxxxxxxxx>.
>
> /Magnus
>
Thank you!
Just one more question regarding *ksoftirqd*-load. This paper seems to talk about the mentioned issue: http://vger.kernel.org/lpc_net2018_talks/lpc18_paper_af_xdp_perf-v2.pdf that *ksoftirqd* is producing high load. A stated solution to this problem would be *busy polling*. I am not sure if *busy polling* is something I have to assign via *setsockopt* in my userspace program or if this already taken care of by libbpf?
Nevertheless, I tried it like this:
$ cat /boot/config-5.4.0-4-amd64 | grep "CONFIG_NET_RX_BUSY_POLL"
CONFIG_NET_RX_BUSY_POLL=y
$ sysctl net.core.busy_poll=50
and in my user-space program:
int busy_poll_usec = 50;
if(setsockopt(xsk_socket__fd(xsk_sockets[i]->xsk), SOL_SOCKET, SO_BUSY_POLL, (char *)&busy_poll_usec, sizeof(busy_poll_usec)) < 0) {
fprintf(stderr, "Failed to set `SO_BUSY_POLL`: %s\n", strerror(errno));
break;
}
/* some code inbetween */
while(!global_exit) {
const int ret = poll(fds, fd_count, 2);
if (ret <= 0) {
continue;
}
for(int i = 0; i < fd_count; i++) {
struct xsk_socket_info *socket = xsk_sockets[i];
if(atomic_exchange(&socket->stats_sync.lock, 1) == 0) {
handle_receive_packets(socket);
atomic_fetch_xor(&socket->stats_sync.lock, 1); /* release socket-lock */
}
}
}
but this only has the effect that my userspace program is now at 50% load (previously around 13%) with *ksoftirqd* still running at 100% (and even worse pps).
Is this the expected effect on what I did?
> > > /Magnus
> > >
> > > >
> > > > > > Just to make sure: Those 650k packets are arriving on the same RX-Queue (even though this NIC has multiple RX-Queues I want to test maximum bandwith for a single RX-Queue).
> > > > > >
> > > > > > > > I didn't observe a dramatic change as I've hoped to. Are there some other ways to reduce interrupt load (user-space application and ksoftirq are already running on different CPUs)?
> > > > > > >
> > > > > > > The need_wakeup flag has a big impact when you run the softirq and the
> > > > > > > application thread on the same core. When using two cores for this, it
> > > > > > > has less of an impact.
> > > > > > >
> > > > > > > /Magnus
> > > > > > >
> > > > > > > > NIC: Mellanox Technologies MT27800
> > > > > > > >
> > > > > > > > Best regards
> > > > > > > >
> > > > > > > > Max
> > > > > > > >
> > > > > >
> > > >
|