Move "tee" building down to interpreter::set_logging_proc

This patch gets rid of this hack in mi_set_logging:

      /* The tee created already is based on gdb_stdout, which for MI
	 is a console and so we end up in an infinite loop of console
	 writing to ui_file writing to console etc.  So discard the
	 existing tee (it hasn't been used yet, and MI won't ever use
	 it), and create one based on raw_stdout instead.  */

By pushing down responsibility for the tee creation to the
interpreter.  I.e., pushing the CLI bits out of handle_redirections
down to the CLI interpreter's set_logging_proc method.

This fixes a few leaks that I spotted, and then confirmed with
"valgrind --leak-check=full":

[...]
  ==21429== 56 (32 direct, 24 indirect) bytes in 1 blocks are definitely lost in loss record 30,243 of 34,980
  ==21429==    at 0x4C29216: operator new(unsigned long) (vg_replace_malloc.c:334)
  ==21429==    by 0x62D9A9: mi_set_logging(interp*, int, ui_file*, ui_file*) (mi-interp.c:1395)
  ==21429==    by 0x810B8A: current_interp_set_logging(int, ui_file*, ui_file*) (interps.c:360)
  ==21429==    by 0x61C537: handle_redirections(int) (cli-logging.c:162)
  ==21429==    by 0x61C6EC: set_logging_on(char*, int) (cli-logging.c:190)
  ==21429==    by 0x6163BE: do_cfunc(cmd_list_element*, char*, int) (cli-decode.c:105)
  ==21429==    by 0x6193C1: cmd_func(cmd_list_element*, char*, int) (cli-decode.c:1913)
  ==21429==    by 0x8DB790: execute_command(char*, int) (top.c:674)
  ==21429==    by 0x632AE6: mi_execute_cli_command(char const*, int, char const*) (mi-main.c:2343)
  ==21429==    by 0x6329BA: mi_cmd_execute(mi_parse*) (mi-main.c:2306)
  ==21429==    by 0x631E19: captured_mi_execute_command(ui_out*, mi_parse*) (mi-main.c:1998)
  ==21429==    by 0x632389: mi_execute_command(char const*, int) (mi-main.c:2163)
  ==21429==
[...]
  ==26635== 24 bytes in 1 blocks are definitely lost in loss record 20,740 of 34,995
  ==26635==    at 0x4C29216: operator new(unsigned long) (vg_replace_malloc.c:334)
  ==26635==    by 0x61C355: handle_redirections(int) (cli-logging.c:131)
  ==26635==    by 0x61C6EC: set_logging_on(char*, int) (cli-logging.c:190)
  ==26635==    by 0x6163BE: do_cfunc(cmd_list_element*, char*, int) (cli-decode.c:105)
  ==26635==    by 0x6193C1: cmd_func(cmd_list_element*, char*, int) (cli-decode.c:1913)
  ==26635==    by 0x8DB7BC: execute_command(char*, int) (top.c:674)
  ==26635==    by 0x7B9132: command_handler(char*) (event-top.c:590)
  ==26635==    by 0x7B94F7: command_line_handler(char*) (event-top.c:780)
  ==26635==    by 0x7B8ABB: gdb_rl_callback_handler(char*) (event-top.c:213)
  ==26635==    by 0x933CE9: rl_callback_read_char (callback.c:220)
  ==26635==    by 0x7B89ED: gdb_rl_callback_read_char_wrapper_noexcept() (event-top.c:175)
  ==26635==    by 0x7B8A49: gdb_rl_callback_read_char_wrapper(void*) (event-top.c:192)

One is fixed by transfering ownership of the log file to the tee.  In
pseudo-patch, since the code was moved at the same time:

 -     out = new tee_file (curr_output, false, logfile.get (), false);
 +     out = new tee_file (curr_output, false, logfile.get (), true);

The other is this bit in mi_set_logging:

    else
      {
 +      delete mi->raw_stdout;

I tried to split the leak fixes to a smaller preparatory patch, but
that was difficult exactly because of the tee hack in
handle_redirections -> mi_set_logging.

gdb/ChangeLog:
2017-02-02  Pedro Alves  <palves@redhat.com>

	* cli/cli-interp.c (struct saved_output_files, saved_output):
	Moved from cli/cli-logging.c.
	(cli_set_logging): New function.
	(cli_interp_procs): Install cli_set_logging.
	* cli/cli-interp.h (make_logging_output, cli_set_logging):
	Declare.
	* cli/cli-logging.c (struct saved_output_files, saved_output):
	Moved to cli/cli-interp.c.
	(pop_output_files): Don't save outputs here.
	(make_logging_output): New function.
	(handle_redirections): Don't build tee nor save previous outputs
	here.
	* interps.c (current_interp_set_logging): Change prototype.
	Assume there's always a set_logging_proc method installed.
	* interps.h (interp_set_logging_ftype): Change prototype.
	(current_interp_set_logging): Change prototype and adjust comment.
	* mi/mi-interp.c (mi_set_logging): Change protototype.  Adjust to
	use make_logging_output.
	* tui/tui-interp.c (tui_interp_procs): Install cli_set_logging.

1 files changed, 10 insertions, 20 deletions

diff --git a/gdb/mi/mi-interp.c b/gdb/mi/mi-interp.c
index f167a53..aa76989 100644
--- a/gdb/mi/mi-interp.c
+++ b/gdb/mi/mi-interp.c
@@ -1373,33 +1373,25 @@ mi_ui_out (struct interp *interp)
 /* Do MI-specific logging actions; save raw_stdout, and change all
    the consoles to use the supplied ui-file(s).  */
 
-static int
-mi_set_logging (struct interp *interp, int start_log,
-		struct ui_file *out, struct ui_file *logfile)
+static void
+mi_set_logging (struct interp *interp,
+	        ui_file_up logfile, bool logging_redirect)
 {
   struct mi_interp *mi = (struct mi_interp *) interp_data (interp);
 
-  if (!mi)
-    return 0;
+  gdb_assert (mi != NULL);
 
-  if (start_log)
+  if (logfile != NULL)
     {
-      /* The tee created already is based on gdb_stdout, which for MI
-	 is a console and so we end up in an infinite loop of console
-	 writing to ui_file writing to console etc.  So discard the
-	 existing tee (it hasn't been used yet, and MI won't ever use
-	 it), and create one based on raw_stdout instead.  */
-      if (logfile)
-	{
-	  delete out;
-	  out = new tee_file (mi->raw_stdout, false, logfile, false);
-	}
-
       mi->saved_raw_stdout = mi->raw_stdout;
-      mi->raw_stdout = out;
+      mi->raw_stdout = make_logging_output (mi->raw_stdout,
+					    std::move (logfile),
+					    logging_redirect);
+
     }
   else
     {
+      delete mi->raw_stdout;
       mi->raw_stdout = mi->saved_raw_stdout;
       mi->saved_raw_stdout = NULL;
     }
@@ -1409,8 +1401,6 @@ mi_set_logging (struct interp *interp, int start_log,
   mi->log->set_raw (mi->raw_stdout);
   mi->targ->set_raw (mi->raw_stdout);
   mi->event_channel->set_raw (mi->raw_stdout);
-
-  return 1;
 }
 
 /* The MI interpreter's vtable.  */