[dpdk-dev] ConnectX-4/mlx5 crashes around rxq_cqe_comp_en?

[Slava Ovsiienko]

Hi, Yasuhiro

Could you, please, create the ticket in Bugzilla.dpdk.org to store the details?
The Rx CQE compression can be disabled with specifying "rxq_cqe_comp_en=0".

WBR, Slava

> -----Original Message-----
> From: dev <dev-bounces at dpdk.org> On Behalf Of Yasuhiro Ohara
> Sent: Friday, July 19, 2019 5:20
> To: dev at dpdk.org
> Subject: Re: [dpdk-dev] ConnectX-4/mlx5 crashes around rxq_cqe_comp_en?
> 
> 
> The same goes with DPDK-19.05 too.
> 
> When crash happens,
> mcqe_n == t_pkt->data_len == 124.
> 
> struct rte_mbuf **elts (which seems to be prepared somewhere) looks like
> it's supposed to contain valid mbufs, but (when under a significant load?) it
> doesn't.
> 
> (gdb) p/x (void*[124])elts[0]
> $31 = {0x1d0bd0d80, 0x1d1cfef80, 0x1d28f6a40, 0x1d22eb100,
> 0x1d195a8c0,
>   0x1d2137200, 0x1d1eb5540, 0x1d1d0fec0, 0x1d28ecf40, 0x1d19b1bc0,
>   0x1cec8a200, 0x1d02e2980, 0x1d085cdc0, 0x1d04e8e00, 0x1ccb4e140,
>   0x1d1e17e80, 0x1d17a1c40, 0x1d14a6e00, 0x1d2871700, 0x1d20b6c40,
>   0x1d29831c0, 0x1d04941c0, 0x1d0921080, 0x1d070ea40, 0x1d148ea80,
>   0x1cee100c0, 0x1d1a47e40, 0x1d0ee6600, 0x1d02f1200, 0x1d24bc100,
>   0x1d1e84e40, 0x1d1e1f2c0, 0x1d28b7ac0, 0x1d2195940, 0x1d21bc540,
>   0x1d228f080, 0x1d1026100, 0x1d285e100, 0x1d211c7c0, 0x1d2128980,
>   0x1d1787200, 0x1d170e080, 0x1d1e0e380, 0x1ce638500, 0x1d21a6880,
>   0x1d20d8ac0, 0x1d25e8600, 0x1d2377880, 0x1d0e13ac0, 0x1c0c07100,
>   0x1c0c07100, 0x1c0c07100, 0x1c0c07100, 0x0, 0x0, 0x0, 0x0,
> 0x7ffff7ff487c,
>   0x1c0c06f00, 0x1c0c08b00, 0x0, 0x0, 0x7ffff7ff207c, 0x1, 0x1480,
>   0x140000000, 0x100000000, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
>   0x40000000ffffffff, 0x4000000000000001, 0xcd64010000000002,
>   0x0 <repeats 46 times>}
> 
> (gdb) p elts[48]
> $38 = (struct rte_mbuf *) 0x1d0e13ac0
> (gdb) p elts[49]
> $39 = (struct rte_mbuf *) 0x1c0c07100
> (gdb) p elts[50]
> $40 = (struct rte_mbuf *) 0x1c0c07100
> (gdb) p elts[51]
> $41 = (struct rte_mbuf *) 0x1c0c07100
> (gdb) p elts[52]
> $42 = (struct rte_mbuf *) 0x1c0c07100
> (gdb) p elts[53]
> $43 = (struct rte_mbuf *) 0x0
> 
> Any thoughts?
> 
> regards,
> Yasu
> 
> From: Yasuhiro Ohara <yasu at nttv6.jp>
> Subject: [dpdk-dev] ConnectX-4/mlx5 crashes around rxq_cqe_comp_en?
> Date: Sat, 13 Jul 2019 01:38:53 +0900 (JST)
> Message-ID: <20190713.013853.751044529514409504.yasu at nttv6.jp>
> 
> >
> > Hi,
> >
> > I get a crash when I put a significant amount of load on
> > ConnectX-4/mlx5, i.e., 50Gbps for 100GbE port.
> >
> > Thread 22 "lcore-slave-19" received signal SIGSEGV, Segmentation fault.
> > [Switching to Thread 0x7fffe77ee700 (LWP 33519)]
> > 0x0000555555f010a3 in _mm_storeu_si128 (__B=..., __P=0x10)
> >     at /usr/lib/gcc/x86_64-linux-gnu/7/include/emmintrin.h:721
> > 721       *__P = __B;
> > (gdb) bt
> > #0  0x0000555555f010a3 in _mm_storeu_si128 (__B=..., __P=0x10)
> >     at /usr/lib/gcc/x86_64-linux-gnu/7/include/emmintrin.h:721
> > #1  rxq_cq_decompress_v (rxq=0x22c910ccc0, cq=0x22c8fd1800,
> elts=0x22c910d240)
> >     at
> > /usr/local/dpdk-stable-18.11.2/drivers/net/mlx5/mlx5_rxtx_vec_sse.h:42
> > 1
> > #2  0x0000555555f04b42 in rxq_burst_v (rxq=0x22c910ccc0,
> pkts=0x7fffe77eba40,
> >     pkts_n=32, err=0x7fffe77dc978)
> >     at
> > /usr/local/dpdk-stable-18.11.2/drivers/net/mlx5/mlx5_rxtx_vec_sse.h:95
> > 6
> > #3  0x0000555555f055ea in mlx5_rx_burst_vec (dpdk_rxq=0x22c910ccc0,
> >     pkts=0x7fffe77eba40, pkts_n=32)
> >     at
> > /usr/local/dpdk-stable-18.11.2/drivers/net/mlx5/mlx5_rxtx_vec.c:238
> > #4  0x0000555555632772 in rte_eth_rx_burst (port_id=4, queue_id=5,
> >     rx_pkts=0x7fffe77eba40, nb_pkts=32)
> >     at
> > /usr/local/dpdk-18.11/x86_64-native-linuxapp-gcc/include/rte_ethdev.h:
> > 3879
> >
> > My environments are:
> >
> > Ubuntu 18.04.2 LTS 4.15.0-50-generic
> > MLNX_OFED_LINUX-4.5-1.0.1.0-ubuntu18.04-x86_64
> > fw_ver: 12.17.2020
> > vendor_id: 0x02c9
> > vendor_part_id: 4115
> > hw_ver: 0x0
> > board_id: LNR3270110033
> > DPDK 18.11.2
> >
> > It looks like the CQE compression is the crashing place.
> >
> > dpdk-stable-18.11.2/drivers/net/mlx5/mlx5_rxtx_vec_sse.h:956
> > 953         /* Decompress the last CQE if compressed. */
> > 954         if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n)
> {
> > 955                 assert(comp_idx == (nocmp_n %
> MLX5_VPMD_DESCS_PER_LOOP));
> > 956                 rxq_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
> >
> > And I'm wondering how I can disable rxq_cqe_comp_en devargs.
> >
> >
> <https://eur03.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdoc
> > .dpdk.org%2Fguides-
> 18.02%2Fnics%2Fmlx5.html&data=02%7C01%7Cviaches
> >
> lavo%40mellanox.com%7C346e0b13aba945bce40808d70befa6e9%7Ca6529
> 71c7d2e4
> >
> d9ba6a4d149256f461b%7C0%7C0%7C636990996240384705&sdata=id
> %2FJY%2BM
> > PhzsUBmFP9YoKtFkcu%2FbJtO6Ntb1QghmnSdQ%3D&reserved=0>
> > 22.5.3. Run-time configuration
> > rxq_cqe_comp_en parameter [int]
> >
> > Any information or guesses are appreciated.
> >
> > Best regards,
> > Yasu
> >