> From: Chengwen Feng <fengchengwen@huawei.com>
> Sent: Thursday, August 1, 2024 12:11 PM
> To: thomas@monjalon.net <thomas@monjalon.net>; dev@dpdk.org <dev@dpdk.org>
> Cc: Van Haaren, Harry <harry.van.haaren@intel.com>; wangchenxingyu@huawei.com <wangchenxingyu@huawei.com>
> Subject: [PATCH] examples/eventdev: fix segment fault with generic pipeline
>
> There was a segmentation fault when executing eventdev_pipeline with
> command [1] with ConnectX-5 NIC card:
>
> 0x000000000079208c in rte_eth_tx_buffer (tx_pkt=0x16f8ed300, buffer=0x100, queue_id=11, port_id=0) at ../lib/ethdev/rte_ethdev.h:6636
> txa_service_tx (txa=0x17b19d080, ev=0xffffffffe500, n=4) at ../lib/eventdev/rte_event_eth_tx_adapter.c:631
> 0x0000000000792234 in txa_service_func (args=0x17b19d080) at ../lib/eventdev/rte_event_eth_tx_adapter.c:666
> 0x00000000008b0784 in service_runner_do_callback (s=0x17fffe100, cs=0x17ffb5f80, service_idx=2) at ../lib/eal/common/rte_service.c:405
> 0x00000000008b0ad8 in service_run (i=2, cs=0x17ffb5f80, service_mask=18446744073709551615, s=0x17fffe100, serialize_mt_unsafe=0)
> at ../lib/eal/common/rte_service.c:441
> 0x00000000008b0c68 in rte_service_run_iter_on_app_lcore (id=2, serialize_mt_unsafe=0) at ../lib/eal/common/rte_service.c:477
> 0x000000000057bcc4 in schedule_devices (lcore_id=0) at ../examples/eventdev_pipeline/pipeline_common.h:138
> 0x000000000057ca94 in worker_generic_burst (arg=0x17b131e80) at ../examples/eventdev_pipeline/pipeline_worker_generic.c:83
> 0x00000000005794a8 in main (argc=11, argv=0xfffffffff470) at ../examples/eventdev_pipeline/main.c:449
>
> The root cause is that the queue_id (11) is invalid, the queue_id comes
> from mbuf.hash.txadapter.txq which may pre-write by NIC driver when
> receiving packets (e.g. pre-write mbuf.hash.fdir.hi field).
Good bug report, thanks for the detailed info on hash.fdir.hi union-ed with txadapter fields.
I don't have the specific HW to test, so code review only.
I don't recall the TXQ quantities etc (been a number of years since I worked on this code...!)
so I'll +CC Pavan who reworked the logic around generic workers & eventdev stages, and might recall?
> Because this example only enabled one ethdev queue, so fixes it by reset
> txq to zero in the first worker stage.
>
> [1] dpdk-eventdev_pipeline -l 0-48 --vdev event_sw0 -- -r1 -t1 -e1 -w ff0 -s5 -n0 -c32 -W1000 -D
> When launch eventdev_pipeline with command [1], event_sw
>
> Fixes: 81fb40f95c82 ("examples/eventdev: add generic worker pipeline")
> Cc: stable@dpdk.org
>
> Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
> Reported-by: Chenxingyu Wang <wangchenxingyu@huawei.com>
Generally the change looks fine - I'll wait a few days for Pavan's input, and otherwise review & Ack assuming no issues found.
Thanks for the patch! -Harry