[PATCH] tracecmd: --poll option is not passed down to guest

commit e7ffcbda9d62855ec30dc36ef5bebcb669321073 added the
--poll option to avoid waking up the trace buffer reader
from an isolated CPU:

    trace-cmd: Add option to poll trace buffers

    Waiting for data to be available on the trace ring-buffers may trigger
    IPIs. This might generate unacceptable trace noise when debugging low
    latency or real time systems. So introduce the poll option. When
    enabled, it forces trace-cmd to use O_NONBLOCK. The drawback to using it
    is that traces will be extracted by busy waiting, which will
    unnecessarily hog the CPUs, so only use when really needed.

    Link: https://lore.kernel.org/linux-trace-devel/[email protected]

    Signed-off-by: Nicolas Saenz Julienne <[email protected]>
    Signed-off-by: Steven Rostedt (VMware) <[email protected]>

However the --poll option is not passed down to the guest, which can cause
the wakeup (and associated latency) in the guest:

84221.535329: kvm_entry:             vcpu 1 rip 0xfff0
84221.535341: write_msr:            6e0, value 97374f0241f8
84221.535343: sched_switch:         cyclictest:6563 [4] S ==> swapper/1:0 [120]
84221.535343: ipi_send_cpu:         cpu=1 callsite=irq_work_queue+0x2f callback=rb_wake_up_waiters+0x0
84221.535344: write_msr:            83f, value f6
84221.535345: irq_work_entry:       vector=246

Signed-off-by: Marcelo Tosatti <[email protected]>


diff --git a/tracecmd/trace-record.c b/tracecmd/trace-record.c
index 4e1157c9..ac42d428 100644
--- a/tracecmd/trace-record.c
+++ b/tracecmd/trace-record.c
@@ -6220,9 +6220,16 @@ static void add_arg(struct buffer_instance *instance,
 {
 	char *ptr, *arg;
 	int i, ret;
+	int long_opt;
+
+	long_opt = 0;
+	for (i = 0; long_options[i].name; i++) {
+		if (long_options[i].val == c)
+			long_opt = 1;
+	}
 
 	/* Short or long arg */
-	if (!(c & 0x80)) {
+	if (long_opt == 0) {
 		ptr = strchr(opts, c);
 		if (!ptr)
 			return; /* Not found? */