> -----Original Message-----
> From: Stephen Hemminger <[email protected]>
> Sent: Wednesday, April 29, 2026 11:12 PM
> To: Zaiyu Wang <[email protected]>; Zaiyu Wang
> <[email protected]>
> Cc: [email protected]; [email protected]; Jiawen Wu <[email protected]>;
> Ferruh Yigit <[email protected]>; [email protected]; [email protected]; Jiawen
> Wu <[email protected]>; Ferruh Yigit <[email protected]>
> Subject: Re: [PATCH v2 11/20] net/txgbe: fix traffic class priority
> configuration
>
> On Wed, 29 Apr 2026 18:25:05 +0800
> Zaiyu Wang <[email protected]> wrote:
>
> > After applying the following testpmd command, 802.1Q packets with
> > specific priorities were not properly directed to the corresponding traffic
> > classes:
> > port config 0 dcb vt off 4 pfc off
> >
> > The old driver had two issues:
> > 1. The hardware uses a 4-bit mapping register per traffic class for
> > priority-to-TC mapping, but the driver incorrectly configured it
> > as 3 bits.
> > 2. The DCB TX configuration mistakenly wrote to the RX register.
> >
> > Fix both issues, ensuring that tc-prio mapping works as expected.
> >
> > Fixes: 8bdc7882f376 ("net/txgbe: support DCB")
> > Cc: [email protected]
> >
> > Signed-off-by: Zaiyu Wang <[email protected]>
> > ---
>
> AI review spotted issues:
>
> Error: The fix is incomplete. The same priority-to-TC layout is also written
> in
> drivers/net/txgbe/txgbe_rxtx.c around line 3380 in txgbe_vmdq_dcb_configure():
> for (i = 0; i < RTE_ETH_DCB_NUM_USER_PRIORITIES; i++)
> /*
> * mapping is done with 3 bits per priority,
> * so shift by i*3 each time
> */
> queue_mapping |= ((cfg->dcb_tc[i] & 0x07) << (i * 3));
>
> wr32(hw, TXGBE_RPUP2TC, queue_mapping);
> This writes the same TXGBE_RPUP2TC register (0x019008) the patch is fixing,
> using a hardcoded i*3 shift and a comment claiming 3 bits per priority. After
> this
> patch, txgbe_dcb_config_rx_arbiter_raptor() encodes priorities at 4-bit
> positions
> (via the updated TXGBE_RPUP2TC_UP_SHIFT=4), but txgbe_vmdq_dcb_configure()
> still uses 3-bit positions. The VMDQ+DCB path will produce a wrong tc-prio
> mapping. txgbe_vmdq_dcb_configure() needs the same fix (use
> TXGBE_RPUP2TC_UP_SHIFT and update the comment).
> Warning: TXGBE_DCBUP2TC_DEC and TXGBE_DCBUP2TC_MAP are now
> inconsistent. After this patch:
> #define TXGBE_DCBUP2TC_MAP(tc, v) LS(v, 4 * (tc), 0x7)
> #define TXGBE_DCBUP2TC_DEC(tc, r) RS(r, 3 * (tc), 0x7)
> DEC will not decode what MAP encoded for any tc > 0. TXGBE_DCBUP2TC_DEC
> isn't currently used in tree, so this is not a runtime bug, but it's a latent
> trap. Either
> update DEC to also use 4*(tc), or remove DEC since it has no callers.
>
Hi Stephen,
Thank you for the thorough AI‑assisted review. I really appreciate it.
I have addressed all the issues you raised, and an updated patch set will be
sent shortly.
