diff options
Diffstat (limited to 'results/classifier/gemma3:12b/socket/764')
| -rw-r--r-- | results/classifier/gemma3:12b/socket/764 | 46 |
1 files changed, 46 insertions, 0 deletions
diff --git a/results/classifier/gemma3:12b/socket/764 b/results/classifier/gemma3:12b/socket/764 new file mode 100644 index 00000000..b4ba6672 --- /dev/null +++ b/results/classifier/gemma3:12b/socket/764 @@ -0,0 +1,46 @@ + +qemu-system-x86 crash (reason: use after free in socket_reconnect_timeout when reconnecting vhost-user dev) +Description of problem: +(gdb) bt<br/> +#0 0x00007f205976b78b in raise () from /usr/lib64/libc.so.6<br/> +#1 0x00007f205976cab1 in abort () from /usr/lib64/libc.so.6<br/> +#2 0x00007f205976404a in ?? () from /usr/lib64/libc.so.6<br/> +#3 0x00007f20597640c2 in __assert_fail () from /usr/lib64/libc.so.6<br/> +#4 0x00007f20594ea556 in **qemu_mutex_lock_impl**(mutex=<optimized out>, file=<optimized out>, line=<optimized out>)<br/> +#5 0x00007f205957a4ef in **socket_reconnect_timeout** (opaque=<optimized out>)<br/> +#6 0x00007f205993b68d in ?? () from /usr/lib64/libglib-2.0.so.0<br/> +#7 0x00007f205993aba4 in g_main_context_dispatch () from /usr/lib64/libglib-2.0.so.0<br/> +#8 0x00007f20594e5d49 in glib_pollfds_poll () at /usr/src/debug/qemu-4.1.0-666.x86_64/util/main-loop.c:218<br/> +#9 0x00007f20594e5dc2 in os_host_main_loop_wait (timeout=<optimized out>)<br/> +#10 0x00007f20594e5f5d in main_loop_wait (nonblocking=nonblocking@entry=0)<br/> +... ...<br/> +#14 0x0000560919e13180 in main (argc=80, argv=0x7ffebc1d0598, envp=0x7ffebc1d0820)<br/> + +at the moment, chr had be free by hot unplug vhost-user dev<br/> + +I think the bug cause reason as following:<br/> +1. when vhost-user dev is connecting state, io-task-worker thread will try call tcp_chr_connect_client_async <br/> + again and again to reconnect.<br/> +2. if reconnect fail, io-task-worker thread will switch to main-thread to handle error, and main-thread will <br/> +call qemu_chr_socket_restart_timer again to reconnect again. <br/> + +3. But, if a hot unplug operation insert to main-thread before io-task-worker switch to main-thread,<br/> + the qemu_chr_socket_restart_timer->socket_reconnect_timeout process will use the released chardev and <br/> + trigger qemu crash + +in short, the primary cause of this bug is io-task-worker reconnect process and <br/> +main-thread hot unplug vhost-user-dev process in a race.<br/> +Steps to reproduce: +1. in qio_task_thread_worker func, add sleep in the following position: <br/> +  task->thread->completion = g_idle_source_new(); <br/> +  g_source_set_callback(task->thread->completion,<br/> + qio_task_thread_result, task, NULL);<br/> +  **sleep(8);**<br/> +  g_source_attach(task->thread->completion,<br/> + task->thread->context);<br/> +  g_source_unref(task->thread->completion); <br/> +2. kill spdk proces or dpdk process, qemu will reconnect to the disconnected vhost-user dev of spdk or dpdk <br/> +3. hot unplug the disconnected vhost-user dev when reconnect logic goto upper sleep position <br/> +4. qemu_chr_socket_restart_timer will use the chr after free, and trigger qemu crash +Additional information: + |