Re: known vboxgetty/isdn issue in 2.6.35.3?

[Arnd Bergmann <arnd@arndb.de>]

On Saturday 06 November 2010, Udo van den Heuvel wrote:
> ==> /proc/16758/stack <==
> [<c118731a>] tty_unthrottle+0x13/0x3a
> [<c1185f4e>] reset_buffer_flags+0xd4/0xd9
> [<c1185f60>] n_tty_flush_buffer+0xd/0x63
> [<c11888a7>] tty_ldisc_flush+0x1f/0x34
> [<c120797c>] isdn_tty_modem_result+0x342/0x37c
> [<c1209c5e>] isdn_tty_modem_hup+0x76/0x176
> [<c120a66b>] isdn_tty_change_speed+0xa2/0xd4
> [<c120a6da>] isdn_tty_set_termios+0x3d/0x5a
> [<c118789b>] set_termios+0x318/0x397
> [<c1187b41>] tty_mode_ioctl+0x178/0x2db
> [<c1187de8>] n_tty_ioctl_helper+0x144/0x154
> [<c11854d9>] n_tty_ioctl+0x97/0xa0
> [<c1183ccd>] tty_ioctl+0x699/0x6d3
> [<c108557c>] vfs_ioctl+0x27/0x91
> [<c1085ade>] do_vfs_ioctl+0x44b/0x47f
> [<c1085b53>] sys_ioctl+0x41/0x61
> [<c10027d3>] sysenter_do_call+0x12/0x2d
> [<ffffffff>] 0xffffffff

Ok, so it seems that set_termios calls change_termios, which holds the
termios_mutex while calling isdn_tty_set_termios. This ends up
calling tty_unthrottle, which tries to take the same mutex, and that fails.

As far as I can tell, the problem got introduced in this commit:

commit 38db89799bdf11625a831c5af33938dcb11908b6
Author: Alan Cox <alan@linux.intel.com>
Date:   Thu Jun 11 12:44:17 2009 +0100
Subject: tty: throttling race fix

However, a fix has recently gone into the kernel that should fix it. As
far as I can tell, 2.6.36 and later should not call tty_ldisc_flush
from isdn_tty_modem_result any more, which avoids the original problem.

Please apply the patch below to see if it helps. We should probably
have it in 2.6.32-stable and 2.6.35-stable, if there is another one.

	Arnd
---
>From bc10f96757bd6ab3721510df8defa8f21c32f974 Mon Sep 17 00:00:00 2001
From: Tilman Schmidt <tilman@imap.cc>
Date: Mon, 5 Jul 2010 14:18:27 +0000
Subject: [PATCH] isdn: avoid calling tty_ldisc_flush() in atomic context

Remove the call to tty_ldisc_flush() from the RESULT_NO_CARRIER
branch of isdn_tty_modem_result(), as already proposed in commit
00409bb045887ec5e7b9e351bc080c38ab6bfd33.
This avoids a "sleeping function called from invalid context" BUG
when the hardware driver calls the statcallb() callback with
command==ISDN_STAT_DHUP in atomic context, which in turn calls
isdn_tty_modem_result(RESULT_NO_CARRIER, ~), and from there,
tty_ldisc_flush() which may sleep.

Signed-off-by: Tilman Schmidt <tilman@imap.cc>
Signed-off-by: David S. Miller <davem@davemloft.net>

diff --git a/drivers/isdn/i4l/isdn_tty.c b/drivers/isdn/i4l/isdn_tty.c
index fc8454d..51dc60d 100644
--- a/drivers/isdn/i4l/isdn_tty.c
+++ b/drivers/isdn/i4l/isdn_tty.c
@@ -2636,12 +2636,6 @@ isdn_tty_modem_result(int code, modem_info * info)
 		if ((info->flags & ISDN_ASYNC_CLOSING) || (!info->tty)) {
 			return;
 		}
-#ifdef CONFIG_ISDN_AUDIO
-		if ( !info->vonline )
-			tty_ldisc_flush(info->tty);
-#else
-		tty_ldisc_flush(info->tty);
-#endif
 		if ((info->flags & ISDN_ASYNC_CHECK_CD) &&
 		    (!((info->flags & ISDN_ASYNC_CALLOUT_ACTIVE) &&
 		       (info->flags & ISDN_ASYNC_CALLOUT_NOHUP)))) {