[PATCH] app/test-pmd: add explicit parameter for fixed link speeds
Bruce Richardson
bruce.richardson at intel.com
Wed Nov 4 14:39:20 CET 2026
On Wed, Nov 04, 2026 at 01:29:49PM +0100, Stephen Hemminger wrote:
> On Tue, 3 Nov 2026 15:50:33 +0000
> Bruce Richardson <bruce.richardson at intel.com> wrote:
>
> > When a particular speed was specified for a port in testpmd, the flags
> > passed to the underlying API call always included the
> > RTE_ETH_LINK_SPEED_FIXED flag. This flag is not a mandatory part of the
> > bitmap when setting speeds, and is used to disable autonegotiation.
> > It's possible to have a link report only a single speed while
> > still have negotiation enabled, so testpmd needs to be extended to
> > support setting a speed both with and without the flag.
> >
> > This patch does so by adding an optional "fixed" keyword after the speed
> > specified by the user. The presence or absense of the keyword adjusts
> > the presence or absense of the FIXED speed bit in the ethdev
> > configuration structure "rte_eth_conf.link_speeds" bitmap.
> >
> > Signed-off-by: Bruce Richardson <bruce.richardson at intel.com>
> > ---
>
> In general it looked ok to me. But I would like testpmd and ethtool have same syntax
> to follow the principle of "least surprise" for users.
>
> The AI review with its issues and that encouragement was:
>
> Review: [PATCH] app/test-pmd: add explicit parameter for fixed link
> speeds (patchwork 172601)
>
> Applies to main (26.11-rc2), builds with -Dwerror=true.
>
> Warning (should fix)
> --------------------
>
> 1. Syntax should follow ethtool rather than invent "fixed".
>
> ethtool has autoneg as a separate on/off setting:
>
> ethtool -s eth0 speed 10000 duplex full autoneg off
>
> With autoneg on, a given speed/duplex limits what gets advertised
> (the kernel's ethnl_auto_linkmodes()). That is the same as this
> patch's default, so the semantics are right; only the keyword and
> its position need to change:
>
> port config (port_id|all) speed (...|auto) duplex (half|full|auto)
> [autoneg (on|off)]
>
> autoneg off -> set RTE_ETH_LINK_SPEED_FIXED
> autoneg on -> clear it
> absent -> leave autoneg alone. ethtool keeps the current
> setting, so keep the FIXED bit already in
> ports[pid].dev_conf.link_speeds. The default
> dev_conf is autoneg, so a fresh port behaves as
> posted.
> speed auto + autoneg off -> error, as now.
>
> Putting the optional tokens last leaves the existing
> "speed N duplex D" prefix alone. Users who know ethtool can guess
> it. Untested sketch for the all-ports case:
>
> struct cmd_config_speed_all {
> ...
> cmdline_fixed_string_t item3; /* "autoneg" */
> cmdline_fixed_string_t value3; /* "on#off" */
> };
>
> static cmdline_parse_token_string_t cmd_config_speed_all_item3 =
> TOKEN_STRING_INITIALIZER(struct cmd_config_speed_all, item3,
> "autoneg");
> static cmdline_parse_token_string_t cmd_config_speed_all_value3 =
> TOKEN_STRING_INITIALIZER(struct cmd_config_speed_all, value3,
> "on#off");
>
> The second cmdline_parse_inst_t still needs the two extra tokens.
> In the handler, res->item3[0] == '\0' means autoneg was not given.
> The cmdline library zeroes the result buffer, so the
> ".data = &cmd_config_speed_all_fixed" self-reference used as a flag
> can go. Same for the per-port instance.
>
Ack, will rework to use the autoneg syntax.
> 2. Default behaviour of an existing command changes silently.
>
> Before this patch "port config 0 speed 10000 duplex full" set
> RTE_ETH_LINK_SPEED_FIXED. After it, the same command leaves autoneg
> enabled and advertises only 10G. Existing scripts and test plans
> (DTS, vendor validation) keep parsing but now do something
> different. Whether the user sees a difference depends on the PMD.
>
> The new default is correct (it matches ethtool), but the release
> note puts it under New Features. Give the behaviour change its own
> sentence, e.g.:
>
> Speed settings no longer disable autonegotiation by default;
> add ``autoneg off`` to force the speed.
>
This becomes moot if we rework the syntax. However, when making the
original change, I had AI do an analysis - only 5 drivers in tree actually
take account of the FIXED flag in turning on/off autoneg. Hopefully, adding
testpmd support will allow that list to increase, because nearly all
drivers just ignore it.
Interestingly, AI also flagged that there is a further set of 5 drivers
which would reject the existing testpmd command as it currently stands
since they reject any speed selection with the FIXED flag set.
> 3. --eth-link-speed is now inconsistent with the runtime command.
>
> parse_link_speed() in parameters.c still starts from
> RTE_ETH_LINK_SPEED_FIXED, and run_app.rst calls it "a forced link
> speed". So the startup option forces the speed and the runtime
> command no longer does. Either add a matching option (e.g.
> --eth-link-autoneg=on|off), or at minimum document it in the speed
> command section:
>
> The ``--eth-link-speed`` command line option always disables
> autonegotiation.
>
If we rework this also becomes largely moot, though the option of a
non-fixed link speed flag is a possibility for the future.
> Info (consider)
> ---------------
>
> 4. testpmd_funcs.rst: the patch leaves three blank lines before
> "port config - queues/descriptors"; one is enough. "Support for
> these settings depends on the driver and hardware." adds nothing
> and can go.
>
> 5. Commit message: "absense" (twice) -> "absence"; "still have
> negotiation enabled" -> "still have autonegotiation enabled".
More information about the dev
mailing list