vlib: add postmortem pcap dispatch trace
Inspired by a real-life conundrum: scenario X involves a vpp crash in
ip4-load-balance because vnet_buffer(b)->ip.adj_index[VLIB_TX] is
(still) set to ~0.
The problem takes most of a day to occur, and we need to see the
broken packet's graph trajectory, metadata, etc. to understand the
problem.
Fix a signed/unsigned ASSERT bug in vlib_get_trace_count().
Rename elog_post_mortem_dump() -> vlib_post_mortem_dump(), add
dispatch trace post-mortem dump.
Add FILTER_FLAG_POST_MORTEM so we can (putatively) capture a ludicrous
number of buffer traces, without actually using more than one dispatch
cycle's worth of memory.
Type: improvement
Signed-off-by: Dave Barach <dave@barachs.net>
Change-Id: If093202ef071df46e290370bd9b33bf6560d30e6
diff --git a/src/vlib/main.c b/src/vlib/main.c
index b868a64..f1de909 100644
--- a/src/vlib/main.c
+++ b/src/vlib/main.c
@@ -734,21 +734,44 @@
}
void
-elog_post_mortem_dump (void)
+vlib_post_mortem_dump (void)
{
vlib_main_t *vm = &vlib_global_main;
elog_main_t *em = &vm->elog_main;
+
u8 *filename;
clib_error_t *error;
- if (!vm->elog_post_mortem_dump)
+ if ((vm->elog_post_mortem_dump + vm->dispatch_pcap_postmortem) == 0)
return;
- filename = format (0, "/tmp/elog_post_mortem.%d%c", getpid (), 0);
- error = elog_write_file (em, (char *) filename, 1 /* flush ring */ );
- if (error)
- clib_error_report (error);
- vec_free (filename);
+ if (vm->dispatch_pcap_postmortem)
+ {
+ clib_error_t *error;
+ pcap_main_t *pm = &vm->dispatch_pcap_main;
+
+ pm->n_packets_to_capture = pm->n_packets_captured;
+ pm->file_name = (char *) format (0, "/tmp/dispatch_post_mortem.%d%c",
+ getpid (), 0);
+ error = pcap_write (pm);
+ pcap_close (pm);
+ if (error)
+ clib_error_report (error);
+ /*
+ * We're in the middle of crashing. Don't try to free the filename.
+ */
+ }
+
+ if (vm->elog_post_mortem_dump)
+ {
+ filename = format (0, "/tmp/elog_post_mortem.%d%c", getpid (), 0);
+ error = elog_write_file (em, (char *) filename, 1 /* flush ring */ );
+ if (error)
+ clib_error_report (error);
+ /*
+ * We're in the middle of crashing. Don't try to free the filename.
+ */
+ }
}
/* *INDENT-OFF* */
@@ -1698,6 +1721,19 @@
return cpu_time_now;
}
+static inline void
+pcap_postmortem_reset (vlib_main_t * vm)
+{
+ pcap_main_t *pm = &vm->dispatch_pcap_main;
+
+ /* Reset the trace buffer and capture count */
+ clib_spinlock_lock_if_init (&pm->lock);
+ vec_reset_length (pm->pcap_data);
+ pm->n_packets_captured = 0;
+ clib_spinlock_unlock_if_init (&pm->lock);
+}
+
+
static_always_inline void
vlib_main_or_worker_loop (vlib_main_t * vm, int is_main)
{
@@ -1799,6 +1835,17 @@
clib_call_callbacks (vm->worker_thread_main_loop_callbacks, vm,
cpu_time_now);
+ /*
+ * When trying to understand aggravating, hard-to-reproduce
+ * bugs: reset / restart the pcap dispatch trace once per
+ * main thread dispatch cycle. All threads share the same
+ * (spinlock-protected) dispatch trace buffer. It might take
+ * a few tries to capture a good post-mortem trace of
+ * a multi-thread issue. Best we can do without a big refactor job.
+ */
+ if (is_main && PREDICT_FALSE (vm->dispatch_pcap_postmortem != 0))
+ pcap_postmortem_reset (vm);
+
/* Process pre-input nodes. */
cpu_time_now = clib_cpu_time_now ();
vec_foreach (n, nm->nodes_by_type[VLIB_NODE_TYPE_PRE_INPUT])
@@ -2303,6 +2350,11 @@
vec_validate (tm->nodes, a->buffer_trace_node_index);
tn = tm->nodes + a->buffer_trace_node_index;
tn->limit += a->buffer_traces_to_capture;
+ if (a->post_mortem)
+ {
+ tm->filter_flag = FILTER_FLAG_POST_MORTEM;
+ tm->filter_count = ~0;
+ }
tm->trace_enable = 1;
}));
/* *INDENT-ON* */
@@ -2328,14 +2380,23 @@
pm->n_packets_captured = 0;
pm->packet_type = PCAP_PACKET_TYPE_vpp;
pm->n_packets_to_capture = a->packets_to_capture;
+ vm->dispatch_pcap_postmortem = a->post_mortem;
/* *INDENT-OFF* */
- foreach_vlib_main (({this_vlib_main->dispatch_pcap_enable = 1;}));
+ foreach_vlib_main (({ this_vlib_main->dispatch_pcap_enable = 1;}));
/* *INDENT-ON* */
}
else
{
/* *INDENT-OFF* */
- foreach_vlib_main (({this_vlib_main->dispatch_pcap_enable = 0;}));
+ foreach_vlib_main ((
+ {
+ this_vlib_main->dispatch_pcap_enable = 0;
+ this_vlib_main->dispatch_pcap_postmortem = 0;
+ tm = &this_vlib_main->trace_main;
+ tm->filter_flag = 0;
+ tm->filter_count = 0;
+ tm->trace_enable = 0;
+ }));
/* *INDENT-ON* */
vec_reset_length (vm->dispatch_buffer_trace_nodes);
if (pm->n_packets_captured)
@@ -2373,6 +2434,7 @@
int rv;
int enable = 0;
int status = 0;
+ int post_mortem = 0;
u32 node_index = ~0, buffer_traces_to_capture = 100;
/* Get a line of input. */
@@ -2402,6 +2464,8 @@
unformat_vlib_node, vm, &node_index,
&buffer_traces_to_capture))
;
+ else if (unformat (line_input, "post-mortem %=", &post_mortem, 1))
+ ;
else
{
return clib_error_return (0, "unknown input `%U'",
@@ -2418,6 +2482,7 @@
a->packets_to_capture = max;
a->buffer_trace_node_index = node_index;
a->buffer_traces_to_capture = buffer_traces_to_capture;
+ a->post_mortem = post_mortem;
rv = vlib_pcap_dispatch_trace_configure (a);
@@ -2498,6 +2563,9 @@
* @cliexstart{vppctl pcap dispatch trace off}
* captured 21 pkts...
* saved to /tmp/dispatchTrace.pcap...
+ * Example of how to start a post-mortem dispatch trace:
+ * pcap dispatch trace on max 20000 buffer-trace
+ * dpdk-input 3000000000 post-mortem
* @cliexend
?*/
/* *INDENT-OFF* */
@@ -2505,7 +2573,7 @@
.path = "pcap dispatch trace",
.short_help =
"pcap dispatch trace [on|off] [max <nn>] [file <name>] [status]\n"
- " [buffer-trace <input-node-name> <nn>]",
+ " [buffer-trace <input-node-name> <nn>][post-mortem]",
.function = dispatch_trace_command_fn,
};
/* *INDENT-ON* */