This patchset add supports for TX zero-copy between guest and host
kernel through vhost. It significantly reduces CPU utilization on the
local host on which the guest is located (It reduced about 50% CPU usage
for single stream test on the host, while 4K message size BW has
increased about 50%). The patchset is based on previous submission and
comments from the community regarding when/how to handle guest kernel
buffers to be released. This is the simplest approach I can think of
after comparing with several other solutions.
This patchset has integrated V3 review comments from community:
1. Add more comments on how to use device ZEROCOPY flag;
2. Change device ZEROCOPY to available bit 31
3. Fix skb header linear allocation when virtio_net GSO is not enabled
It has integrated V4 review comments from MST and Sridhar:
1. In vhost, using socket poll wake up for outstanding DMAs
2. Add detailed comments for vhost_zerocopy_signal_used call
3. Add sleep in vhost shutting down instead of busy-wait for outstanding
DMAs.
4. Copy small packets, don't do zero-copy callback in mavtap, mark it's
DMA done in vhost
5. change zerocopy to bool in macvtap.
It integrates V5 review comments from MST and
Michał Mirosław <[email protected]>
1. Prevent userspace apps from holding skb userspace buffers by copying
userspace buffers to kernel in skb_clone, skb_copy, pskb_copy,
pskb_expand_head.
2. It is also used HIGHDMA, SG feature bits to enable ZEROCOPY to remove
the dependency of a new feature bit, we can add it later when new
feature bit is available.
This patchset includes:
1/4: Add a new sock zero-copy flag, SOCK_ZEROCOPY;
2/4: Add a new struct skb_ubuf_info in skb_share_info for userspace
buffers release callback when lower device DMA has done for that skb,
which is the last reference count gone; Or whenever skb_clone, skb_copy,
pskb_copy, pskb_expand_head get call from tcpdump, filtering, these userspace
buffers will be copied into kernel ... we don't want userspace apps to hold
userspace buffers too long.
3/4: Add vhost zero-copy callback in vhost when skb last refcnt is gone;
add vhost_zerocopy_signal_used to notify guest to release TX skb
buffers.
4/4: Add macvtap zero-copy in lower device when sending packet is
greater than 256 bytes.
The patchset is built against most recent net-next linux 2.6.39-rc7. It
has passed netperf/netserver multiple streams stress test, tcpdump
suspended test, dynamically SG change test.
Single TCP_STREAM 120 secs test results over ixgbe 10Gb NIC results:
Message BW(Gb/s)qemu-kvm (NumCPU)vhost-net(NumCPU) PerfTop irq/s
4K 7408.57 92.1% 22.6% 1229
4K(Orig)4913.17 118.1% 84.1% 2086
8K 9129.90 89.3% 23.3% 1141
8K(Orig)7094.55 115.9% 84.7% 2157
16K 9178.81 89.1% 23.3% 1139
16K(Orig)8927.1 118.7% 83.4% 2262
64K 9171.43 88.4% 24.9% 1253
64K(Orig)9085.85 115.9% 82.4% 2229
For message size less or equal than 2K, there is a known KVM guest TX
overrun issue. With this zero-copy patch, the issue becomes more severe,
guest io_exits has tripled than before, so the performance is not good.
Once the TX overrun problem has been addressed, I will retest the small
message size performance.
drivers/net/macvtap.c | 132 ++++++++++++++++++++++++++++++++++++---
drivers/vhost/net.c | 44 +++++++++++++-
drivers/vhost/vhost.c | 49 +++++++++++++++
drivers/vhost/vhost.h | 13 ++++
include/linux/netdevice.h | 10 +++
include/linux/skbuff.h | 26 ++++++++
include/net/sock.h | 1 +
net/core/skbuff.c | 81 ++++++++++++++++++++++++-
8 files changed, 345 insertions(+), 17 deletions(-)