Fixes an obvious bug in cdc-acm by avoiding a recursive lock on
acm_start_wb()'s error path. Should apply towards 2.6.27 stable and
2.6.28.
=============================================
[ INFO: possible recursive locking detected ]
2.6.27-2-pae #109
---------------------------------------------
python/31449 is trying to acquire lock:
 (&acm->write_lock){++..}, at: [<
f89a0348>] acm_start_wb+0x5c/0x7b [cdc_acm]
but task is already holding lock:
 (&acm->write_lock){++..}, at: [<
f89a04fb>] acm_tty_write+0xe1/0x167 [cdc_acm]
other info that might help us debug this:
2 locks held by python/31449:
 #0:  (&tty->atomic_write_lock){--..}, at: [<
c0260fae>] tty_write_lock+0x14/0x3b
 #1:  (&acm->write_lock){++..}, at: [<
f89a04fb>] acm_tty_write+0xe1/0x167 [cdc_acm]
stack backtrace:
Pid: 31449, comm: python Not tainted 2.6.27-2-pae #109
 [<
c030f42f>] ? printk+0xf/0x18
 [<
c0149f33>] __lock_acquire+0xc7b/0x1316
 [<
c014a63e>] lock_acquire+0x70/0x97
 [<
f89a0348>] ? acm_start_wb+0x5c/0x7b [cdc_acm]
 [<
c0312109>] _spin_lock_irqsave+0x37/0x47
 [<
f89a0348>] ? acm_start_wb+0x5c/0x7b [cdc_acm]
 [<
f89a0348>] acm_start_wb+0x5c/0x7b [cdc_acm]
 [<
f89a055d>] acm_tty_write+0x143/0x167 [cdc_acm]
 [<
c0262a98>] write_chan+0x1cd/0x297
 [<
c012527e>] ? default_wake_function+0x0/0xd
 [<
c026111e>] tty_write+0x149/0x1b9
 [<
c02628cb>] ? write_chan+0x0/0x297
 [<
c01912c5>] ? rw_verify_area+0x76/0x98
 [<
c0260fd5>] ? tty_write+0x0/0x1b9
 [<
c01919ba>] vfs_write+0x8c/0x136
 [<
c0191afd>] sys_write+0x3b/0x60
 [<
c0103beb>] sysenter_do_call+0x12/0x3f
 =======================
Signed-off-by: Brandon Philips <bphilips@suse.de>
Cc: Oliver Neukum <oliver@neukum.org>
Cc: stable <stable@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
 }
 
 /*
- * Finish write.
+ * Finish write. Caller must hold acm->write_lock
  */
 static void acm_write_done(struct acm *acm, struct acm_wb *wb)
 {
-       unsigned long flags;
-
-       spin_lock_irqsave(&acm->write_lock, flags);
        wb->use = 0;
        acm->transmitting--;
-       spin_unlock_irqrestore(&acm->write_lock, flags);
 }
 
 /*
 {
        struct acm_wb *wb = urb->context;
        struct acm *acm = wb->instance;
+       unsigned long flags;
 
        if (verbose || urb->status
                        || (urb->actual_length != urb->transfer_buffer_length))
                        urb->transfer_buffer_length,
                        urb->status);
 
+       spin_lock_irqsave(&acm->write_lock, flags);
        acm_write_done(acm, wb);
+       spin_unlock_irqrestore(&acm->write_lock, flags);
        if (ACM_READY(acm))
                schedule_work(&acm->work);
        else