Re: [PATCH v2 0/7] Add vhost-net kernel support to pasta

[Eugenio Perez Martin <eperezma@redhat.com>]

On Fri, Oct 16, 2026 at 10:33 PM Ammar Yasser <aerosound161@gmail.com> wrote:
>
> This work is based on the original v2 RFC by Eugenio Perez:
>
> Link:
> https://archives.passt.top/passt-dev/20250709174748.3514693-1-eperezma@redhat.com/
>
> It is the first finished patch series for vhost-kernel support after a
> third RFC was submitted:
>
> Link:
> https://archives.passt.top/passt-dev/20260802132155.870796-1-aerosound161@gmail.com/
>
> vhost-net is a kernel device that allows reading and writing packets to a
> tap device using virtqueues instead of read(2), write(2), and their
> vectored counterparts. Frames are handed over through a shared descriptor
> ring, so several of them can be passed to the kernel at once rather than
> issuing a syscall per frame.
>
> The v3 RFC didn't yield any performance improvements because, to retain
> correctness, sends between pasta and the vhost-kernel thread had to be
> synchronous. This patch series fixes this problem through the changes
> introduced in the last commit, titled:
>
> "tap/tcp: Replace tcp_payload_used with a ring buffer style index"
>
> The changes reliably yield a roughly 2x speedup on the TX path (container
> to host).
>
> Numbers were obtained through iperf3 with the following methodology and
> results:
>
> Pasta command:
> ./pasta -f --runas 0:0 --map-guest-addr 192.168.0.5
> --vhost-kernel on --config-net 2729785
>
> Host:
> iperf3 -s1J
>
> Container:
> iperf3 -c 192.168.0.5 -t 10 -P4 -l 1M -w 32M -i1
>
> With --vhost-kernel on:
>
> [SUM]   0.00-10.00  sec  1.36 GBytes  1.17 Gbits/sec  3643 sender
> [SUM]   0.00-10.00  sec  1.16 GBytes   999 Mbits/sec        receiver
>
> With --vhost-kernel off:
>
> [SUM]   0.00-10.00  sec   700 MBytes   587 Mbits/sec  3556 sender
> [SUM]   0.00-10.10  sec   495 MBytes   411 Mbits/sec        receiver
>

Sounds good :). Sorry I didn't have the time to review the series
properly, but I suspect you can use VIRTIO F IN ORDER to both:
* Increase the performance even more.
* Make the code simpler as pasta fits (or it used to fit) very nicely
with it. It would have make *my* code simpler at least, so I'm
suspecting it will make this series simpler too.

Does that help?

Thanks!

> v2:
> - removed all use of epoll_ctl in vhost_setup_eventfds in favor of epoll_add
> - prevent all functions that register memory regions into the vhost_memory
>   object from overflowing
> - remove the unions with flexible array members in vring_avail_pasta and
>   vring_used_pasta in favor of hardcoding the ring array to have VHOST_NDESCS
>   entries to avoid having to ignore Wpedantic
> - remove the && operator with TCP_CURR_IDX when computing no_csum
> - make desc_id in tx_reap a uint32_t
> - don't make vhost-kernel a required argument
>
> Ammar Yasser (7):
>   tap: Move the tap_hdr file to a separate file
>   conf: Add context fields, epoll types and --vhost-kernel flag to pasta
>   virtio: Add the pasta vhost-net interface and implementation
>   tap: Implement the pasta vhost-net from-guest path
>   tap, tcp, udp: Prepare the to-guest path for vhost-net
>   tap: Implement the pasta vhost-net to-guest path
>   tap/tcp: Replace tcp_payload_used with a ring buffer style index
>
>  Makefile     |   2 +-
>  conf.c       |  19 +++
>  epoll_type.h |   4 +
>  passt.c      |   9 ++
>  passt.h      |  79 ++++++++++-
>  tap.c        | 368 +++++++++++++++++++++++++++++++++++++++++++++------
>  tap.h        |  47 ++-----
>  tap_hdr.h    |  23 ++++
>  tcp_buf.c    | 149 ++++++++++++++++-----
>  tcp_buf.h    |   4 +
>  udp.c        |  47 +++++--
>  udp.h        |   3 +
>  util.c       |  18 +++
>  util.h       |   2 +
>  vhost.c      | 295 +++++++++++++++++++++++++++++++++++++++++
>  vhost.h      |  91 +++++++++++++
>  16 files changed, 1036 insertions(+), 124 deletions(-)
>  create mode 100644 tap_hdr.h
>  create mode 100644 vhost.c
>  create mode 100644 vhost.h
>
> --
> 2.39.5 (Apple Git-154)
>