memif thread race condition on memif.disconnect()

[Stephen Hemminger]

On Wed, 11 Oct 2023 19:57:56 +0000
"Bly, Mike" <mbly at ciena.com> wrote:

> Hello,
> 
> We have run into a timing issue between threads when using the memif
> interface type and need some guidance.
> 
> Our application has a DPDK based process operating (among other
> things) a memif server interface. The problem is exposed when this
> memif interface receives a memif.disconnect message from the remote
> client, while in the middle of an rte_eth_rx_burst() on this same
> memif interface. As the IRQ message handling is on its own thread as
> compared to the DPDK worker thread doing the rx_burst, this resulted
> in a crash. The backtraces for which have been shared below. How does
> one ensure there are guard rails in place to gracefully exit the
> rx-burst when a disconnect occurs? Or, how do we properly modify the
> code such that  we defer responding to the disconnect CB after the
> rx-burst operation has completed?
> 
> We are utilizing DPDK 21.11.2. I have diff'd dpdks-stable:22.11.3 in
> ./drivers/net/memif, but I do not see anything obvious that would
> address this. I did a similar diff for dpdk:23.07, but do not see
> anything obvious there either.
> 
> -Mike
> 
> (gdb) thread 1
> [Switching to thread 1 (Thread 0x7f17e2813600 (LWP 470))]
> #0  0x00007f17e374d225 in eth_memif_rx (queue=0x1189023b00,
> bufs=0x7f17e28100e8, nb_pkts=32) at
> ../git/drivers/net/memif/rte_eth_memif.c:338 338
> last_slot = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE); (gdb) bt
> #0  0x00007f17e374d225 in eth_memif_rx (queue=0x1189023b00,
> bufs=0x7f17e28100e8, nb_pkts=32) at
> ../git/drivers/net/memif/rte_eth_memif.c:338 #1  0x000000000047e6fb
> in rte_eth_rx_burst (nb_pkts=32, rx_pkts=0x7f17e28100e8, queue_id=0,
> port_id=<optimized out>) at /usr/include/rte_ethdev.h:5368 #2
> pmd_main_loop () at ../git/swfw/api/src/swfwPmd.c:1086 #3
> 0x000000000047f309 in pmd_launch_one_lcore (dummy=<optimized out>) at
> ../git/my_process.c:1157 #4  0x00007f17f7070e7c in eal_thread_loop
> (arg=<optimized out>) at ../git/lib/eal/linux/eal_thread.c:146 #5
> 0x00007f17f4c3da72 in start_thread (arg=<optimized out>) at
> pthread_create.c:442 #6  0x00007f17f4cbf930 in clone3 () at
> ../sysdeps/unix/sysv/linux/x86_64/clone3.S:81 (gdb) l 333
> ring_size = 1 << mq->log2_ring_size; 334             mask = ring_size
> - 1; 335 336             if (type == MEMIF_RING_C2S) { 337
>          cur_slot = mq->last_head; 338                     last_slot
> = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE); 339             }
> else { 340                     cur_slot = mq->last_tail; 341
>            last_slot = __atomic_load_n(&ring->tail,
> __ATOMIC_ACQUIRE); 342             } (gdb) p ring->head Cannot access
> memory at address 0x7f17d8e58006
> 
> (gdb) thread 19
> [Switching to thread 19 (Thread 0x7f17f0804600 (LWP 468))]
> #0  0x00007f17f4caf97b in __GI___close (fd=494) at
> ../sysdeps/unix/sysv/linux/close.c:27 27        return SYSCALL_CANCEL
> (close, fd); (gdb) bt
> #0  0x00007f17f4caf97b in __GI___close (fd=494) at
> ../sysdeps/unix/sysv/linux/close.c:27 #1  0x00007f17e374f01f in
> memif_free_regions (dev=dev at entry=0x7f17f727f000
> <rte_eth_devices+99072>) at
> ../git/drivers/net/memif/rte_eth_memif.c:882 #2  0x00007f17e37475d0
> in memif_disconnect (dev=0x7f17f727f000 <rte_eth_devices+99072>) at
> ../git/drivers/net/memif/memif_socket.c:623 #3  0x00007f17f7091bd2 in
> eal_intr_process_interrupts (nfds=<optimized out>, events=<optimized
> out>) at ../git/lib/eal/linux/eal_interrupts.c:1026 #4
> out>eal_intr_handle_interrupts (totalfds=<optimized out>, pfd=20) at
> out>../git/lib/eal/linux/eal_interrupts.c:1100 #5
> out>eal_intr_thread_main (arg=<optimized out>) at
> out>../git/lib/eal/linux/eal_interrupts.c:1172 #6  0x00007f17f4c3da72
> out>in start_thread (arg=<optimized out>) at pthread_create.c:442 #7
> out>0x00007f17f4cbf930 in clone3 () at
> out>../sysdeps/unix/sysv/linux/x86_64/clone3.S:81
> 

I don't think memif maintainer has been very active.
One possibility would be the memif driver support removal event
interrupt. This would require driver and application change.