[Bug 208409] [PATCH] igb and ALTQ

[bugzilla-noreply at freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=208409

ncrogers at gmail.com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ncrogers at gmail.com

--- Comment #10 from ncrogers at gmail.com ---
ALTQ has never worked with the multiqueue drivers. em(4) interfaces work "out
of the box" when ALTQ is included in the kernel, because the default em(4)
behavior has multiqueue disabled. With igb(4) and ixgbe(4), you MUST enable the
legacy path in the driver. Prior to this patch, the only way to sensibly do
this was to manually edit the driver and add an IGB_LEGACY_TX define.

It sounds like there is an issue with the legacy single-queue path, perhaps
only when ALTQ is not actually used in the PF configuration? FWIW I support
many PF routers with igb(4) interfaces having ALTQ enabled (not just compiled
in the kernel), including the IGB_LEGACY_TX path, and have not had any
problems, albeit I am still running 10.1-RELEASE.

This was a much-needed patch from the perspective of someone that supports many
PF+ALTQ installations and has had to add IGB_LEGACY_TX manually for years now
(since the 8.x days I believe). However it appears that there are many people
that have haphazardly added ALTQ support to their kernels without actually
using it, and that is definitely not a case where you would want to downgrade
to the single-queue behavior.

Somehow this patch in question made it through even though the issue has been
brought up many times before over the past few years. At one point there was
discussion of adding some kind of kernel tunable to enable the IGB/IXGBE legacy
transmit paths, but it did not gain any committer support. Perhaps this is the
way to go as it would have prevented the problem where people are now
unexpectedly running the single-queue transmit path because ALTQ is in their
kernel.

The following PR had discussed the kernel tunable option.
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=194197

-- 
You are receiving this mail because:
You are the assignee for the bug.