When changing so that the client passes its stdout and stderr as well as

stdin up to the server, I forgot one essential point - the tmux server
could now be both the producer and consumer. This happens when tmux is
run inside tmux, as well as when piping tmux commands together.

So, using stdio(3) was a bad idea - if sufficient data was written, this
could block in write(2). When that happened and the server was both
producer and consumer, it deadlocks.

Change to use libevent bufferevents for the client stdin, stdout and
stderr instead. This is trivial enough for output but requires a
callback mechanism to trigger when stdin is finished.

This relies on the underlying polling mechanism for libevent to work
with whatever devices to which the user could redirect stdin, stdout or
stderr, hence the change to use poll(2) over kqueue(2) for tmux.

1 files changed, 1 insertions, 4 deletions

diff --git a/cmd-run-shell.c b/cmd-run-shell.c
index f9d1e237..7bfb8916 100644
--- a/cmd-run-shell.c
+++ b/cmd-run-shell.c
@@ -131,13 +131,10 @@ cmd_run_shell_free(void *data)
 {
 	struct cmd_run_shell_data	*cdata = data;
 	struct cmd_ctx			*ctx = &cdata->ctx;
-	struct msg_exit_data		 exitdata;
 
 	if (ctx->cmdclient != NULL) {
 		ctx->cmdclient->references--;
-		exitdata.retcode = ctx->cmdclient->retcode;
-		server_write_client(
-		    ctx->cmdclient, MSG_EXIT, &exitdata, sizeof exitdata);
+		ctx->cmdclient->flags |= CLIENT_EXIT;
 	}
 	if (ctx->curclient != NULL)
 		ctx->curclient->references--;