ovn-trace: need to read latest data when using daemon mode

[gujun4990]

When logical flow changed in SB, the result that ovn-trace evaluate using daemon mode is previous. Because the data from db just is read once when ovn-trace startup. This situation is not adapted for daemon mode. So we need to keep latest data from SB when ovn-trace use daemon mode.

In addition, print the unixctl server path used to be recorded refer from ovn-nbctl and ovn-sbctl.

Signed-off-by: Jun Gu [email protected]

[gujun4990]

@dceara I do some optimizations for ovn-trace, plz help to review this patch. Thanks!
In addition, I am considering that the already_read variable should be removed.

• After ovn-trace has connected to SB, run read_db once and after trace to return for non-daemon mode.
• For daemon mode, need to read latest data from SB and will run read_db when SB data changed.

So I think the already_read variable can be removed.

-    bool already_read = false;
     for (;;) {
         ovsdb_idl_run(ovnsb_idl);
         unixctl_server_run(server);
@@ -150,10 +156,7 @@ main(int argc, char *argv[])
         }
 
         if (ovsdb_idl_has_ever_connected(ovnsb_idl)) {
-            if (!already_read) {
-                already_read = true;
-                read_db();
-            }
+            read_db();
 
             daemonize_complete();
             if (!get_detach()) {

[dceara]

I didn't review this thoroughly but checking what read_db() does I see we'll be doing this at every iteration:

static void
read_db(void)
{
    read_datapaths();
    read_ports();
    read_mcgroups();
    read_address_sets();
    read_port_groups();
    read_gen_opts();
    read_flows();
    read_mac_bindings();
    read_fdbs();
}

That doesn't seem OK because we never cleanup any of the maps we populate there. E.g., read_datapaths() will reinsert the same datapaths over and over again in the global struct hmap datapaths variable.

[gujun4990]

I observe that the memory keeps growing if using the above code. I will re-optimize the code according to your advice.

[dceara]

We need to incrementally process changes (or clear global data at every iteration) if we want to process DB updates.

[gujun4990]

We need to incrementally process changes (or clear global data at every iteration) if we want to process DB updates.

I will try to optimize it as you suggested.

[gujun4990]

I try to clear global data at every iteration and the code is as follows:

static void
ovntrace_fdbs_destroy(struct hmap *fdbs)
{
    struct ovntrace_fdb *fdb;
    HMAP_FOR_EACH_POP (fdb, node, fdbs) {
        free(fdb);
    }
    hmap_destroy(fdbs);
}

static void
ovntrace_mac_bindings_destroy(struct hmap *bindings)
{
    struct ovntrace_mac_binding *binding;
    HMAP_FOR_EACH_POP (binding, node, bindings) {
        free(binding);
    }
    hmap_destroy(bindings);
}

static void
ovntrace_flows_destroy(struct ovntrace_flow **flows, size_t n_flows)
{
    for (size_t i = 0; i < n_flows; i++) {
        struct ovntrace_flow *flow = flows[i];
        free(flow->stage_name);
        free(flow->source);
        expr_destroy(flow->match);
        ovnacts_free(flow->ovnacts, flow->ovnacts_len);
        free(flow);
    }
    free(flows);
}

static void
ovntrace_mcgroups_destroy(struct ovs_list *mcgroups)
{
    struct ovntrace_mcgroup *mcgroup;
    LIST_FOR_EACH_POP (mcgroup, list_node, mcgroups) {
        free(mcgroup->name);
        free(mcgroup->ports);
        free(mcgroup);
    }
}

static void
ovntrace_datapaths_destroy(void)
{
    struct ovntrace_datapath *dp;
    HMAP_FOR_EACH_POP (dp, sb_uuid_node, &datapaths) {
        ovntrace_mcgroups_destroy(&dp->mcgroups);
        ovntrace_mac_bindings_destroy(&dp->mac_bindings);
        ovntrace_fdbs_destroy(&dp->fdbs);
        ovntrace_flows_destroy(dp->flows, dp->n_flows);

        free(dp->name);
        free(dp->name2);
        free(dp->friendly_name);
        free(dp);
    }
    hmap_destroy(&datapaths);
}

static void
ovntrace_ports_destroy(void)
{
    struct shash_node *node, *next;
    SHASH_FOR_EACH_SAFE (node, next, &ports) {
        struct ovntrace_port *port = node->data;
        free(port->name);
        free(port->type);
        free(port->name2);
        free(port->friendly_name);
        free(port);
        shash_delete(&ports, node);
    }
    shash_destroy(&ports);
}

static void
address_sets_destroy(void)
{
    expr_const_sets_destroy(&address_sets);
    shash_destroy(&address_sets);
}

static void
port_groups_destroy(void)
{
    expr_const_sets_destroy(&port_groups);
    shash_destroy(&port_groups);
}

static void
clear_data(void)
{
    static bool first_clear = true;
    if (first_clear) {
        first_clear = false;
    } else {
        ovntrace_datapaths_destroy();
        ovntrace_ports_destroy();
        address_sets_destroy();
        port_groups_destroy();
        dhcp_opts_destroy(&dhcp_opts);
        dhcp_opts_destroy(&dhcpv6_opts);
        nd_ra_opts_destroy(&nd_ra_opts);
        controller_event_opts_destroy(&event_opts);
        expr_symtab_destroy(&symtab);
        shash_destroy(&symtab);
    }
}

But the memory will still continue to grow when run the above code as daemon mode, but slower than before. So I wonder whether there are some tools can be used to analyze this situation. @dceara

[dceara]

But the memory will still continue to grow when run the above code as daemon mode, but slower than before. So I wonder whether there are some tools can be used to analyze this situation. @dceara

We run all our unit tests with AddressSanitizer enabled. You could try that too. You need to build with the sanitizer enabled, e.g.:

ovn/.ci/linux-build.sh

Lines 47 to 48 in 52da241

	COMMON_CFLAGS="${COMMON_CFLAGS} -O1 -fno-omit-frame-pointer -fno-common -g"
	OVN_CFLAGS="${OVN_CFLAGS} -fsanitize=address,undefined"

And then configure address sanitizer to track memory leaks too before starting the ovn-trace daemon, e.g.:

ovn/tests/atlocal.in

Lines 214 to 215 in 52da241

	ASAN_OPTIONS=detect_leaks=1:abort_on_error=true:log_path=sanitizers:$ASAN_OPTIONS
	export ASAN_OPTIONS

When the ovn-trace daemon exits it will write a sanitizers.<pid> file if there were any memory leaks.

Hope this helps!

[gujun4990]

I re-optimize ovn-trace code. The memory will not keep to grow when using daemon mode. In next stage, we plan to use I-P engine to optimize ovn-trace.