We used to hold the mutex of paired virtqueue in
vhost_net_busy_poll(). But this will results an inconsistent lock
order which may cause deadlock if we try to bring back the protection
of device IOTLB with vq mutex that requires to hold mutex of all
virtqueues at the same time.
Fix this simply by switching to use mutex_trylock(), when fail just
skip the busy polling. This can happen when device IOTLB is under
updating which should be rare.
Fixes: commit 78139c94dc ("net: vhost: lock the vqs one by one")
Cc: Tonghao Zhang <xiangxia.m.yue@gmail.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch improves the guest receive performance.
On the handle_tx side, we poll the sock receive queue at the
same time. handle_rx do that in the same way.
We set the poll-us=100us and use the netperf to test throughput
and mean latency. When running the tests, the vhost-net kthread
of that VM, is alway 100% CPU. The commands are shown as below.
Rx performance is greatly improved by this patch. There is not
notable performance change on tx with this series though. This
patch is useful for bi-directional traffic.
netperf -H IP -t TCP_STREAM -l 20 -- -O "THROUGHPUT, THROUGHPUT_UNITS, MEAN_LATENCY"
Topology:
[Host] ->linux bridge -> tap vhost-net ->[Guest]
TCP_STREAM:
* Without the patch: 19842.95 Mbps, 6.50 us mean latency
* With the patch: 37598.20 Mbps, 3.43 us mean latency
Signed-off-by: Tonghao Zhang <xiangxia.m.yue@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Factor out generic busy polling logic and will be
used for in tx path in the next patch. And with the patch,
qemu can set differently the busyloop_timeout for rx queue.
To avoid duplicate codes, introduce the helper functions:
* sock_has_rx_data(changed from sk_has_rx_data)
* vhost_net_busy_poll_try_queue
Signed-off-by: Tonghao Zhang <xiangxia.m.yue@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We accidentally left out this error return so it leads to some use after
free bugs later on.
Fixes: 0a0be13b8f ("vhost_net: batch submitting XDP buffers to underlayer sockets")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch implements XDP batching for vhost_net. The idea is first to
try to do userspace copy and build XDP buff directly in vhost. Instead
of submitting the packet immediately, vhost_net will batch them in an
array and submit every 64 (VHOST_NET_BATCH) packets to the under layer
sockets through msg_control of sendmsg().
When XDP is enabled on the TUN/TAP, TUN/TAP can process XDP inside a
loop without caring GUP thus it can do batch map flushing. When XDP is
not enabled or not supported, the underlayer socket need to build skb
and pass it to network core. The batched packet submission allows us
to do batching like netif_receive_skb_list() in the future.
This saves lots of indirect calls for better cache utilization. For
the case that we can't so batching e.g when sndbuf is limited or
packet size is too large, we will go for usual one packet per
sendmsg() way.
Doing testpmd on various setups gives us:
Test /+pps%
XDP_DROP on TAP /+44.8%
XDP_REDIRECT on TAP /+29%
macvtap (skb) /+26%
Netperf tests shows obvious improvements for small packet transmission:
size/session/+thu%/+normalize%
64/ 1/ +2%/ 0%
64/ 2/ +3%/ +1%
64/ 4/ +7%/ +5%
64/ 8/ +8%/ +6%
256/ 1/ +3%/ 0%
256/ 2/ +10%/ +7%
256/ 4/ +26%/ +22%
256/ 8/ +27%/ +23%
512/ 1/ +3%/ +2%
512/ 2/ +19%/ +14%
512/ 4/ +43%/ +40%
512/ 8/ +45%/ +41%
1024/ 1/ +4%/ 0%
1024/ 2/ +27%/ +21%
1024/ 4/ +38%/ +73%
1024/ 8/ +15%/ +24%
2048/ 1/ +10%/ +7%
2048/ 2/ +16%/ +12%
2048/ 4/ 0%/ +2%
2048/ 8/ 0%/ +2%
4096/ 1/ +36%/ +60%
4096/ 2/ -11%/ -26%
4096/ 4/ 0%/ +14%
4096/ 8/ 0%/ +4%
16384/ 1/ -1%/ +5%
16384/ 2/ 0%/ +2%
16384/ 4/ 0%/ -3%
16384/ 8/ 0%/ +4%
65535/ 1/ 0%/ +10%
65535/ 2/ 0%/ +8%
65535/ 4/ 0%/ +1%
65535/ 8/ 0%/ +3%
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch introduces to a new tun/tap specific msg_control:
#define TUN_MSG_UBUF 1
#define TUN_MSG_PTR 2
struct tun_msg_ctl {
int type;
void *ptr;
};
This allows us to pass different kinds of msg_control through
sendmsg(). The first supported type is ubuf (TUN_MSG_UBUF) which will
be used by the existed vhost_net zerocopy code. The second is XDP
buff, which allows vhost_net to pass XDP buff to TUN. This could be
used to implement accepting an array of XDP buffs from vhost_net in
the following patches.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We use to have message like:
struct vhost_msg {
int type;
union {
struct vhost_iotlb_msg iotlb;
__u8 padding[64];
};
};
Unfortunately, there will be a hole of 32bit in 64bit machine because
of the alignment. This leads a different formats between 32bit API and
64bit API. What's more it will break 32bit program running on 64bit
machine.
So fixing this by introducing a new message type with an explicit
32bit reserved field after type like:
struct vhost_msg_v2 {
__u32 type;
__u32 reserved;
union {
struct vhost_iotlb_msg iotlb;
__u8 padding[64];
};
};
We will have a consistent ABI after switching to use this. To enable
this capability, introduce a new ioctl (VHOST_SET_BAKCEND_FEATURE) for
userspace to enable this feature (VHOST_BACKEND_F_IOTLB_V2).
Fixes: 6b1e6cc785 ("vhost: new device IOTLB API")
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Like commit e2b3b35eb9 ("vhost_net: batch used ring update in rx"),
this patches implements batch used ring update for datacopy TX
(zerocopy has already done some kind of batching).
Testpmd transmission from guest to host (XDP_DROP on tap) shows 25.8%
improvement (from ~3.1Mpps to ~3.9Mpps) on Broadwell i7-5600U CPU @
2.60GHz machine. Netperf TCP tests does not show obvious differences.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
A more generic name which could be used for TX as well.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Instead of mixing zerocopy and datacopy logics, this patch tries to
split datacopy logic out. This results for a more compact code and
ad-hoc optimization could be done on top more easily.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduce tx_can_batch() to determine whether TX could be
batched. This will help to reduce the code duplication in the future.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Factor out logic of getting tx buffer and iov iter
initialization. This will be used for reducing codes duplication in
the future.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduce init_iov_iter() in order to be reused by future patch.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We may run out of avail rx ring descriptor under heavy load but busypoll
did not detect it so busypoll may have exited prematurely. Avoid this by
checking rx ring full during busypoll.
Signed-off-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp>
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We may run handle_rx() while rx work is queued. For example a packet can
push the rx work during the window before handle_rx calls
vhost_net_disable_vq().
In that case busypoll immediately exits due to vhost_has_work()
condition and enables vq again. This can lead to another unnecessary rx
wake-ups, so poll rx work instead of enabling the vq.
Signed-off-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp>
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Under heavy load vhost busypoll may run without suppressing
notification. For example tx zerocopy callback can push tx work while
handle_tx() is running, then busyloop exits due to vhost_has_work()
condition and enables notification but immediately reenters handle_tx()
because the pushed work was tx. In this case handle_tx() tries to
disable notification again, but when using event_idx it by design
cannot. Then busyloop will run without suppressing notification.
Another example is the case where handle_tx() tries to enable
notification but avail idx is advanced so disables it again. This case
also leads to the same situation with event_idx.
The problem is that once we enter this situation busyloop does not work
under heavy load for considerable amount of time, because notification
is likely to happen during busyloop and handle_tx() immediately enables
notification after notification happens. Specifically busyloop detects
notification by vhost_has_work() and then handle_tx() calls
vhost_enable_notify(). Because the detected work was the tx work, it
enters handle_tx(), and enters busyloop without suppression again.
This is likely to be repeated, so with event_idx we are almost not able
to suppress notification in this case.
To fix this, poll the work instead of enabling notification when
busypoll is interrupted by something. IMHO vhost_has_work() is kind of
interruption rather than a signal to completely cancel the busypoll, so
let's run busypoll after the necessary work is done.
Signed-off-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp>
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Sock will be NULL if we pass -1 to vhost_net_set_backend(), but when
we meet errors during ubuf allocation, the code does not check for
NULL before calling sockfd_put(), this will lead NULL
dereferencing. Fixing by checking sock pointer before.
Fixes: bab632d69e ("vhost: vhost TX zero-copy support")
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Filling in the padding slot in the bpf structure as a bug fix in 'ne'
overlapped with actually using that padding area for something in
'net-next'.
Signed-off-by: David S. Miller <davem@davemloft.net>
Jason Wang
Tonghao Zhang
Dan Carpenter
Toshiaki Makita
Kees Cook
David S. Miller