On 02/24/2015 04:31 PM, Yao Qi wrote:
> Pedro Alves <palves@redhat.com> writes:
> 
> Hi Pedro,
> 
>> +	# GDBserver doesn't do inferior I/O through GDB.  But we can
>> +	# talk to the program using GDBserver's tty instead.
>> +	global inferior_spawn_id
>> +	set inferior_spawn_id $server_spawn_id
> 
> Does it still work well if GDBserver is started without tty?  In my
> remote testing, gdbserver is started without tty,
> 
> spawn /usr/bin/ssh -l yao junor1 /gdbserver/aarch64/gdbserver --once :2346 aarch64-linux-gnu/gdb/testsuite/gdb.base/interrupt
> 
> I see the following timeouts:
> 
> Continuing.^M
> Remote debugging from host 10.2.206.34^M
> FAIL: gdb.base/interrupt.exp: process is alive (timeout)
> a^M
> a^M
> FAIL: gdb.base/interrupt.exp: child process ate our char (timeout)

I tried it now, and the problem is that when the gdbserver/inferior is
started without a tty, then stdout is put in unbuffered mode, so the
printf calls don't flush...

Adding:
  setvbuf (stdout, NULL, _IONBF, BUFSIZ);

to interrupt.c fixes it.

This is the same problem we see when testing with Windows (native or
remote) from Cygwin, given in that case stdin/stdout is connected
to a pipe.

Long ago I added the set_unbuffered_mode.c hack to for this.  So enabling
that hack fixes this too then.  See patch below. (I also hacked
native-gdbserver.exp locally to use ssh with no -t to test it).  That would
require boards to set that  gdb,force_unbuffered_mode flag if they need it.

But I'm not sure we want to expose that to boards.  We could also
always enable the hack for gdbserver in gdbserver-support.exp.

Or we could fix the tests themselves to explicitly call setvbuf
if needed and not bother boards at all.  I count only around 20
tests that check gdb,noinferiorio, or use gdb_skip_stdio_test, and
we could fix them incrementally, as they're converted to
use $inferior_spawn_id.  Maybe that's the cleanest.  We can
e.g., add:

 #include "lib/set_unbuffered_mode.c"

at the top of such files, which avoids an explicit call in
"main".   That relies on __attribute__ ((constructor)), but
we could also call an helper shared function that does the
setvbuf from the tests' "main" if we don't want to rely on
that attribute.

Options, options...

> 
> We need to override ${board}_spawn and pass "-t" to ssh.  After this change,
> all interrupt.exp tests pass.  Since the test harness assumes GDBserver
> has tty, probably we should document such requirement somewhere.

That's an option too, but it makes me a bit nervous.  I'm not sure
if we can assume that.

> 
> However, I don't run the whole testsuite with the updated board file
> (with -t option to ssh).
>