Re: stmmac and XDP/ZC issue

[Maciej Fijalkowski <maciej.fijalkowski@intel.com>]

On Wed, Feb 21, 2024 at 10:21:04AM +0100, Kurt Kanzenbach wrote:
> On Wed Feb 21 2024, Kurt Kanzenbach wrote:
> > On Tue Feb 20 2024, Stanislav Fomichev wrote:
> >> On Tue, Feb 20, 2024 at 6:43 AM Maciej Fijalkowski
> >> <maciej.fijalkowski@intel.com> wrote:
> >>>
> >>> On Tue, Feb 20, 2024 at 04:18:54PM +0300, Serge Semin wrote:
> >>> > Hi Kurt
> >>> >
> >>> > On Tue, Feb 20, 2024 at 12:02:25PM +0100, Kurt Kanzenbach wrote:
> >>> > > Hello netdev community,
> >>> > >
> >>> > > after updating to v6.8 kernel I've encountered an issue in the stmmac
> >>> > > driver.
> >>> > >
> >>> > > I have an application which makes use of XDP zero-copy sockets. It works
> >>> > > on v6.7. On v6.8 it results in the stack trace shown below. The program
> >>> > > counter points to:
> >>> > >
> >>> > >  - ./include/net/xdp_sock.h:192 and
> >>> > >  - ./drivers/net/ethernet/stmicro/stmmac/stmmac_main.c:2681
> >>> > >
> >>> > > It seems to be caused by the XDP meta data patches. This one in
> >>> > > particular 1347b419318d ("net: stmmac: Add Tx HWTS support to XDP ZC").
> >>> > >
> >>> > > To reproduce:
> >>> > >
> >>> > >  - Hardware: imx93
> >>> > >  - Run ptp4l/phc2sys
> >>> > >  - Configure Qbv, Rx steering, NAPI threading
> >>> > >  - Run my application using XDP/ZC on queue 1
> >>> > >
> >>> > > Any idea what might be the issue here?
> >>> > >
> >>> > > Thanks,
> >>> > > Kurt
> >>> > >
> >>> > > Stack trace:
> >>> > >
> >>> > > |[  169.248150] imx-dwmac 428a0000.ethernet eth1: configured EST
> >>> > > |[  191.820913] imx-dwmac 428a0000.ethernet eth1: EST: SWOL has been switched
> >>> > > |[  226.039166] imx-dwmac 428a0000.ethernet eth1: entered promiscuous mode
> >>> > > |[  226.203262] imx-dwmac 428a0000.ethernet eth1: Register MEM_TYPE_PAGE_POOL RxQ-0
> >>> > > |[  226.203753] imx-dwmac 428a0000.ethernet eth1: Register MEM_TYPE_PAGE_POOL RxQ-1
> >>> > > |[  226.303337] imx-dwmac 428a0000.ethernet eth1: Register MEM_TYPE_XSK_BUFF_POOL RxQ-1
> >>> > > |[  255.822584] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
> >>> > > |[  255.822602] Mem abort info:
> >>> > > |[  255.822604]   ESR = 0x0000000096000044
> >>> > > |[  255.822608]   EC = 0x25: DABT (current EL), IL = 32 bits
> >>> > > |[  255.822613]   SET = 0, FnV = 0
> >>> > > |[  255.822616]   EA = 0, S1PTW = 0
> >>> > > |[  255.822618]   FSC = 0x04: level 0 translation fault
> >>> > > |[  255.822622] Data abort info:
> >>> > > |[  255.822624]   ISV = 0, ISS = 0x00000044, ISS2 = 0x00000000
> >>> > > |[  255.822627]   CM = 0, WnR = 1, TnD = 0, TagAccess = 0
> >>> > > |[  255.822630]   GCS = 0, Overlay = 0, DirtyBit = 0, Xs = 0
> >>> > > |[  255.822634] user pgtable: 4k pages, 48-bit VAs, pgdp=0000000085fe1000
> >>> > > |[  255.822638] [0000000000000000] pgd=0000000000000000, p4d=0000000000000000
> >>> > > |[  255.822650] Internal error: Oops: 0000000096000044 [#1] PREEMPT_RT SMP
> >>> > > |[  255.822655] Modules linked in:
> >>> > > |[  255.822660] CPU: 0 PID: 751 Comm: napi/eth1-261 Not tainted 6.8.0-rc4-rt4-00100-g9c63d995ca19 #8
> >>> > > |[  255.822666] Hardware name: NXP i.MX93 11X11 EVK board (DT)
> >>> > > |[  255.822669] pstate: 60400009 (nZCv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
> >>> > > |[  255.822674] pc : stmmac_tx_clean.constprop.0+0x848/0xc38
> >>> > > |[  255.822690] lr : stmmac_tx_clean.constprop.0+0x844/0xc38
> >>> > > |[  255.822696] sp : ffff800085ec3bc0
> >>> > > |[  255.822698] x29: ffff800085ec3bc0 x28: ffff000005b609e0 x27: 0000000000000001
> >>> > > |[  255.822706] x26: 0000000000000000 x25: ffff000005b60ae0 x24: 0000000000000001
> >>> > > |[  255.822712] x23: 0000000000000001 x22: ffff000005b649e0 x21: 0000000000000000
> >>> > > |[  255.822719] x20: 0000000000000020 x19: ffff800085291030 x18: 0000000000000000
> >>> > > |[  255.822725] x17: ffff7ffffc51c000 x16: ffff800080000000 x15: 0000000000000008
> >>> > > |[  255.822732] x14: ffff80008369b880 x13: 0000000000000000 x12: 0000000000008507
> >>> > > |[  255.822738] x11: 0000000000000040 x10: 0000000000000a70 x9 : ffff800080e32f84
> >>> > > |[  255.822745] x8 : 0000000000000000 x7 : 0000000000000000 x6 : 0000000000003ff0
> >>> > > |[  255.822751] x5 : 0000000000003c40 x4 : ffff000005b60000 x3 : 0000000000000000
> >>> > > |[  255.822757] x2 : 0000000000000000 x1 : 0000000000000000 x0 : 0000000000000000
> >>> > > |[  255.822764] Call trace:
> >>> > > |[  255.822766]  stmmac_tx_clean.constprop.0+0x848/0xc38
> >>>
> >>> Shouldn't xsk_tx_metadata_complete() be called only when corresponding
> >>> buf_type is STMMAC_TXBUF_T_XSK_TX?
> >>
> >> +1. I'm assuming Serge isn't enabling it explicitly, so none of the
> >> metadata stuff should trigger in this case.
> >
> > The only other user of xsk_tx_metadata_complete() in mlx5 guards it with
> > xp_tx_metadata_enabled(). Seems like that's missing in stmmac?
> 
> Well, the following patch seems to help:
> 
> commit e85ab4b97b4d6e50036435ac9851b876c221f580
> Author: Kurt Kanzenbach <kurt@linutronix.de>
> Date:   Wed Feb 21 08:18:15 2024 +0100
> 
>     net: stmmac: Complete meta data only when enabled
>     
>     Currently using XDP sockets on stmmac results in a kernel crash:
>     
>     |[  255.822584] Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
>     |[...]
>     |[  255.822764] Call trace:
>     |[  255.822766]  stmmac_tx_clean.constprop.0+0x848/0xc38
>     
>     The program counter indicates xsk_tx_metadata_complete(). However, this
>     function shouldn't be called unless metadata is actually enabled.
>     
>     Tested on imx93.
>     
>     Fixes: 1347b419318d ("net: stmmac: Add Tx HWTS support to XDP ZC")
>     Signed-off-by: Kurt Kanzenbach <kurt@linutronix.de>
> 
> diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> index 9df27f03a8cb..77c62b26342d 100644
> --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
> @@ -2678,9 +2678,10 @@ static int stmmac_tx_clean(struct stmmac_priv *priv, int budget, u32 queue,
>                                         .desc = p,
>                                 };
>  
> -                               xsk_tx_metadata_complete(&tx_q->tx_skbuff_dma[entry].xsk_meta,
> -                                                        &stmmac_xsk_tx_metadata_ops,
> -                                                        &tx_compl);
> +                               if (xp_tx_metadata_enabled(tx_q->xsk_pool))

every other usage of tx metadata functions should be wrapped with
xp_tx_metadata_enabled() - can you address other places and send a proper
patch?

> +                                       xsk_tx_metadata_complete(&tx_q->tx_skbuff_dma[entry].xsk_meta,
> +                                                                &stmmac_xsk_tx_metadata_ops,
> +                                                                &tx_compl);
>                         }
>                 }