api: Avoid the usage of the freed registration by the API calls

This issue happens if:
- the API client connects via Unix socket
- the client issues the *_dump API call and immediately disconnects

What happens after is that the API handler keeps sending the *_details
messages, however at some point the write fails, and the socket is
deleted.

The attempt of a use of the registration pointer results in interpreting
the socket as a shared memory socket. This results in a crash, because
the data in this structure then does not make sense, like the below:

|
|Thread 1 "vpp_main" received signal SIGSEGV, Segmentation fault.
|__GI___pthread_mutex_lock (mutex=0x0) at ../nptl/pthread_mutex_lock.c:67
|67      ../nptl/pthread_mutex_lock.c: No such file or directory.
|(gdb) bt
|#0  __GI___pthread_mutex_lock (mutex=0x0) at ../nptl/pthread_mutex_lock.c:67
|#1  0x00007ffff500f957 in svm_queue_lock (q=0x0) at /home/ubuntu/vpp/src/svm/queue.c:101
|#2  svm_queue_add (q=0x0, elem=0x7fffa76c2de0 "\210\365\006\060\001", nowait=0) at /home/ubuntu/vpp/src/svm/queue.c:274
|#3  0x00007ffff6e131e3 in vl_api_send_msg (rp=<optimized out>, elem=<optimized out>) at /home/ubuntu/vpp/src/vlibmemory/api.h:43
|#4  send_sw_interface_details (am=<optimized out>, rp=<optimized out>, swif=0x7fffb957a0bc, interface_name=<optimized out>, context=<optimized out>)
|    at /home/ubuntu/vpp/src/vnet/interface_api.c:353
|#5  0x00007ffff6e0edeb in vl_api_sw_interface_dump_t_handler (mp=<optimized out>) at /home/ubuntu/vpp/src/vnet/interface_api.c:412
|#6  0x00007ffff7daeb48 in msg_handler_internal (am=<optimized out>, the_msg=0x7fffb839a5e0, trace_it=<optimized out>, do_it=1, free_it=0)
|    at /home/ubuntu/vpp/src/vlibapi/api_shared.c:501
|#7  vl_msg_api_socket_handler (the_msg=0x7fffb839a5e0) at /home/ubuntu/vpp/src/vlibapi/api_shared.c:790
|#8  0x00007ffff7d7c608 in vl_socket_process_api_msg (rp=<optimized out>, input_v=0x7fffa76c2de0 "\210\365\006\060\001") at /home/ubuntu/vpp/src/vlibmemory/socket_api.c:212
|#9  0x00007ffff7d89ff1 in vl_api_clnt_process (vm=<optimized out>, node=<optimized out>, f=<optimized out>) at /home/ubuntu/vpp/src/vlibmemory/vlib_api.c:405
|#10 0x00007ffff53bf9a7 in vlib_process_bootstrap (_a=<optimized out>) at /home/ubuntu/vpp/src/vlib/main.c:1490
|#11 0x00007ffff4da0b2c in clib_calljmp () from /home/ayourtch/vpp/build-root/install-vpp-native/vpp/lib/libvppinfra.so.21.06
|#12 0x00007fffa99a4d90 in ?? ()
|#13 0x00007ffff53b6cb2 in vlib_process_startup (vm=0x7ffff56a9880 <vlib_global_main>, p=0x7fffb5d41380, f=0x0) at /home/ubuntu/vpp/src/vlib/main.c:1515
|#14 dispatch_process (vm=0x7ffff56a9880 <vlib_global_main>, p=0x7fffb5d41380, f=0x0, last_time_stamp=<optimized out>) at /home/ubuntu/vpp/src/vlib/main.c:1571
|#15 0x0000000000000000 in ?? ()
|(gdb) frame 3
|#3  0x00007ffff6e131e3 in vl_api_send_msg (rp=<optimized out>, elem=<optimized out>) at /home/ubuntu/vpp/src/vlibmemory/api.h:43
|43            vl_msg_api_send_shmem (rp->vl_input_queue, (u8 *) & elem);
|(gdb) l
|38          {
|39            vl_socket_api_send (rp, elem);
|40          }
|41        else
|42          {
|43            vl_msg_api_send_shmem (rp->vl_input_queue, (u8 *) & elem);
|44          }
|45      }
|46
|47      always_inline int
|(gdb)
|

The approach in this change is to avoid the closing operations "here and
now", but instead mark the the registration as a zombie and place
a forced RPC towards a callback that does the actual cleanup work.

Forced RPC is handled via the API processing loop with barrier sync,
so we are guaranteed not to have any API processing in-process.

Type: fix
Change-Id: I1972d42da620bdb4fd773c83262863c2781d9005
Signed-off-by: Andrew Yourtchenko <ayourtch@gmail.com>

0 files changed, 0 insertions, 0 deletions