[dpdk-dev] [PATCH v3 0/2] TAP RSS eBPF cover letter

Ophir Munk ophirmu at mellanox.com
Wed Jan 10 16:34:21 CET 2018


Hi Jason,
Thanks for your inputs.
Please see inline

> -----Original Message-----
> From: Jason Wang [mailto:jasowang at redhat.com]
> Sent: Wednesday, January 10, 2018 12:11 PM
> To: Ophir Munk <ophirmu at mellanox.com>; dev at dpdk.org; Pascal Mazon
> <pascal.mazon at 6wind.com>
> Cc: Thomas Monjalon <thomas at monjalon.net>; Olga Shern
> <olgas at mellanox.com>
> Subject: Re: [dpdk-dev] [PATCH v3 0/2] TAP RSS eBPF cover letter
> 
> 
> 
> On 2018年01月10日 15:06, Ophir Munk wrote:
> > The patches of TAP RSS eBPF follow the RFC on this issue
> >
> https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdpd
> >
> k.org%2Fdev%2Fpatchwork%2Fpatch%2F31781%2F&data=02%7C01%7Coph
> irmu%40me
> >
> llanox.com%7C8edca4ebac3440cf465c08d55812768b%7Ca652971c7d2e4d9
> ba6a4d1
> >
> 49256f461b%7C0%7C0%7C636511758701461920&sdata=2WBaV4iU9kE5ub
> o2lgL0adDH
> > %2BcTxce6SP4KBevyRcsI%3D&reserved=0
> >
> > v3 changes with respect to v2
> > =============================
> > * Add support for IPv6 RSS in BPF program
> > * Bug fixes
> > * Updated compatibility to kernel versions:
> >    eBPF requires Linux version 4.9 configured with BPF
> > * New license header (SPDX) for newly added files
> >
> > v2 changes with respect to v1
> > =============================
> > * v2 has new commits organization (3 --> 2)
> > * BPF program was revised. It is successfully tested on
> >    IPv4 L3 L4 layers (compatible to mlx4 device)
> > * Licensing: no comments received for using "Dual BSD/GPL"
> >    string during BPF program loading to the kernel.
> >    (v1 and v2 are using the same license strings)
> >    Any comments are welcome.
> > * Compatibility to kernel versions:
> >    eBPF requires Linux version 4.2 configured with BPF. TAP PMD will
> >    successfully compile on systems with old or non-BPF configured kernels.
> >    During compilation time the required Linux headers are searched for.
> >    If they are not present missing definitions are locally added
> >    (tap_autoconf.h).
> >    If the kernel cannot support a BPF operation - at runtime it will
> >    gracefully reject the netlink message (with BPF) sent to it.
> >
> > Commit #1:
> > net/tap: add eBPF instructions to TAP device
> > ===========================================
> >    This commit introduces BPF program (tap_bpf_program.c)
> >    with a classifier and an action sections.
> >    The current implementation calculates RSS hash
> >    over L3 addresses and L4 ports.
> >    BPF program compilation is not part of dpdk compilation.
> >    This commit includes the eBPF machine instructions
> >    in the format of C arrays (tap_bpf_insns.c).
> >
> >    The option to update the BPF program and download new machine
> >    instructions will be described in another commit.
> >
> > Commit #2:
> > net/tap: implement RSS with eBPF classifier and action
> > ======================================================
> > This commit builds and sends netlink messages to the kernel that
> > include BPF programs.
> > There is a single BPF classifier for each TAP queue. Packets marked
> > with an RSS queue will be directed to this queue using a traffic
> > control with "skbedit" action otherwise they will be pipelined to the
> > following rules.
> > There is a single BPF action for each RSS rule (may add more to
> > support IPv6).
> > The action is to calculate Toeplitz hash based on L3 addresses and L4
> > ports, mark the packet with the RSS queue according the resulting
> > hash, then reclassify the packet.
> > Ophir Munk (2):
> >    net/tap: add eBPF instructions to TAP device
> >    net/tap: implement RSS with eBPF classifier and action
> >
> >   drivers/net/tap/Makefile          |   31 +
> >   drivers/net/tap/rte_eth_tap.h     |    9 +-
> >   drivers/net/tap/tap_bpf.h         |   92 ++
> >   drivers/net/tap/tap_bpf_insns.c   | 1905
> +++++++++++++++++++++++++++++++++++++
> >   drivers/net/tap/tap_bpf_program.c |  221 +++++
> >   drivers/net/tap/tap_flow.c        |  635 +++++++++++--
> >   drivers/net/tap/tap_flow.h        |   15 +
> >   drivers/net/tap/tap_rss.h         |   32 +
> >   drivers/net/tap/tap_tcmsgs.h      |    4 +
> >   9 files changed, 2848 insertions(+), 96 deletions(-)
> >   create mode 100644 drivers/net/tap/tap_bpf.h
> >   create mode 100644 drivers/net/tap/tap_bpf_insns.c
> >   create mode 100644 drivers/net/tap/tap_bpf_program.c
> >   create mode 100644 drivers/net/tap/tap_rss.h
> >
> 
> TAP will support eBPF classification directly [1] through eBPF socket filter.
> Compare to tc-bpf, it was more portable for other backends (e.g macvtap).
> 

I agree that classifying directly through eBPF using socket filter is easier than building a set of rules in TC, however there are more concerns that justify using eBPF with TC. 
1. An RSS rule can come in the middle of non-RSS rules. It seems natural to have it as part of TC. How do I know if a packet needs to be treated with RSS or not? (there may be a higher priority rule that sends the packet to a specific queue regardless of RSS).
2. BPF complexity comes mainly in writing BPF programs, compiling them and sending BPF instructions to the kernel. In both mechanisms (socket filter or TC) this complexity is the same.

Please let me have your feedback.

> [1]
> https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgit.
> kernel.org%2Fpub%2Fscm%2Flinux%2Fkernel%2Fgit%2Fdavem%2Fnet-
> next.git%2Fcommit%2Fdrivers%2Fnet%2Ftun.c%3Fid%3D96f84061620c6325
> a2ca9a9a05b410e6461d03c3&data=02%7C01%7Cophirmu%40mellanox.co
> m%7C8edca4ebac3440cf465c08d55812768b%7Ca652971c7d2e4d9ba6a4d1
> 49256f461b%7C0%7C0%7C636511758701461920&sdata=8zPaZlAHYSsFCQw
> XcDLtro5QcrIN5pwKsj%2FUwFJ9q8s%3D&reserved=0
> 
> Thanks


More information about the dev mailing list