Dear Ole,
I have a quick/easy request for the Python language binding. See
pneum_connect(...). We need to make the client input queue size configurable,
with a default size of 32. See below.
Seems like a simple change to vpp_papi.py:connect(...) ->
pneum_wrap.c:wrap_connect(...) -> pneum.c:pneum_connect(...). I hate to make
work for other people, but I’m pretty sure that I’ll foul up the
PyArg_ParseTuple(...) bit.
Why do we care? When run at scale, Wojciech’s test code bumps against its input
queue size.
Vpp deals with the situation reasonably gracefully, but [especially in
single-core cases] the data plane is forced to turf vpp->client messages
[stats, interface event notifications, etc.] to avoid gross packet loss
[especially when an API client dies]. We can’t let the data-plane wait for a
[possibly dead] client to clean up its input queue.
Thanks... Dave
int
pneum_connect (char * name, char * chroot_prefix, pneum_callback_t cb)
{
int rv = 0;
pneum_main_t *pm = &pneum_main;
if (chroot_prefix != NULL)
vl_set_memory_root_path (chroot_prefix);
if ((rv = vl_client_api_map("/vpe-api"))) {
clib_warning ("vl_client_api map rv %d", rv);
return rv;
}
if (vl_client_connect(name, 0, 32) < 0) { /* Python code must be able to set
this parameter */
vl_client_api_unmap();
return (-1);
}
if (cb) {
/* Start the rx queue thread */
rv = pthread_create(&pm->rx_thread_handle, NULL, pneum_rx_thread_fn, 0);
if (rv) {
clib_warning("pthread_create returned %d", rv);
vl_client_api_unmap();
return (-1);
}
pneum_callback = cb;
}
pm->connected_to_vlib = 1;
return (0);
}
smime.p7s
Description: S/MIME cryptographic signature
_______________________________________________ vpp-dev mailing list vpp-dev@lists.fd.io https://lists.fd.io/mailman/listinfo/vpp-dev
