On Tue, Apr 1, 2025 at 12:04 PM Jon Kohler <j...@nutanix.com> wrote: > > In handle_tx_copy, TX batching processes packets below ~PAGE_SIZE and > batches up to 64 messages before calling sock->sendmsg. > > Currently, when there are no more messages on the ring to dequeue, > handle_tx_copy re-enables kicks on the ring *before* firing off the > batch sendmsg. However, sock->sendmsg incurs a non-zero delay, > especially if it needs to wake up a thread (e.g., another vhost worker). > > If the guest submits additional messages immediately after the last ring > check and disablement, it triggers an EPT_MISCONFIG vmexit to attempt to > kick the vhost worker. This may happen while the worker is still > processing the sendmsg, leading to wasteful exit(s). > > This is particularly problematic for single-threaded guest submission > threads, as they must exit, wait for the exit to be processed > (potentially involving a TTWU), and then resume. > > In scenarios like a constant stream of UDP messages, this results in a > sawtooth pattern where the submitter frequently vmexits, and the > vhost-net worker alternates between sleeping and waking. > > A common solution is to configure vhost-net busy polling via userspace > (e.g., qemu poll-us). However, treating the sendmsg as the "busy" > period by keeping kicks disabled during the final sendmsg and > performing one additional ring check afterward provides a significant > performance improvement without any excess busy poll cycles. > > If messages are found in the ring after the final sendmsg, requeue the > TX handler. This ensures fairness for the RX handler and allows > vhost_run_work_list to cond_resched() as needed. > > Test Case > TX VM: taskset -c 2 iperf3 -c rx-ip-here -t 60 -p 5200 -b 0 -u -i 5 > RX VM: taskset -c 2 iperf3 -s -p 5200 -D > 6.12.0, each worker backed by tun interface with IFF_NAPI setup. > Note: TCP side is largely unchanged as that was copy bound > > 6.12.0 unpatched > EPT_MISCONFIG/second: 5411 > Datagrams/second: ~382k > Interval Transfer Bitrate Lost/Total Datagrams > 0.00-30.00 sec 15.5 GBytes 4.43 Gbits/sec 0/11481630 (0%) sender > > 6.12.0 patched > EPT_MISCONFIG/second: 58 (~93x reduction) > Datagrams/second: ~650k (~1.7x increase) > Interval Transfer Bitrate Lost/Total Datagrams > 0.00-30.00 sec 26.4 GBytes 7.55 Gbits/sec 0/19554720 (0%) sender > > Signed-off-by: Jon Kohler <j...@nutanix.com>
Acked-by: Jason Wang <jasow...@redhat.com> Thanks
