Messages in this thread Patch in this message | | | Date | Tue, 24 Sep 2013 14:18:15 -0400 | From | Peter Hurley <> | Subject | Re: [PATCH 1/1] tty: disassociate_ctty() sends the extra SIGCONT |
| |
On 09/22/2013 04:03 PM, Oleg Nesterov wrote: > On 09/21, Peter Hurley wrote: >> >> On 09/21/2013 02:34 PM, Oleg Nesterov wrote: >>> >>> do_each_pid_task(tty->session, PIDTYPE_SID, p) { >>> spin_lock_irq(&p->sighand->siglock); >>> if (p->signal->tty == tty) { >>> p->signal->tty = NULL; >>> /* We defer the dereferences outside fo >>> the tasklist lock */ >>> refs++; >>> } >>> if (!p->signal->leader) { >>> spin_unlock_irq(&p->sighand->siglock); >>> continue; >>> } >>> __group_send_sig_info(SIGHUP, SEND_SIG_PRIV, p); >>> __group_send_sig_info(SIGCONT, SEND_SIG_PRIV, p); >>> put_pid(p->signal->tty_old_pgrp); /* A noop */ >>> spin_lock(&tty->ctrl_lock); >>> tty_pgrp = get_pid(tty->pgrp); >>> >>> I guess this can happen only once, so we could even add WARN_ON(tty_pgrp) >>> before get_pid(). But this look confusing, as if we can do get_pid() >>> multiple times and leak tty->pgrp. >>> >>> if (tty->pgrp) >>> p->signal->tty_old_pgrp = get_pid(tty->pgrp); >>> >>> else? We already did put_pid(tty_old_pgrp), we should clear it. >>> >>> IOW, do you think the patch below makes sense or I missed something? >>> Just curious. >> >> The code block you're referring to only executes once because there is >> only one session leader. > > I understand, and I even mentioned this above.
Ah, ok. I didn't realize the earlier patch was a cleanup attempt and not fixing something.
> My point was, this _looks_ confusing, and the patch I sent makes it > more clean.
I agree that this looks confusing, but I'm not sure I agree that your earlier patch makes it cleaner; maybe a code comment stating that the block only executes once for the session leader would be more appropriate.
Also, I put the get_pid() in the siglock critical section to prevent unsafe racing between hangup and ioctl(TIOCSCTTY).
> And what about ->tty_old_pgrp? I still think that at least the patch > below makes sense. If tty->pgrp == NULL is not possible here (I do > not know), then why do we check?
tty->pgrp can be NULL here if the session leader is dropping the controlling terminal association via no_tty(). But in this case ->tty_old_pgrp will also be NULL.
This race should probably be eliminated by claiming the tty_lock() in no_tty(), so that it doesn't race with __tty_hangup() at all.
[NB: The other possibility, a second hangup, is no longer possible.]
> Otherwise ->tty_old_pgrp != NULL looks certainly wrong after put_pid().
I agree this looks fairly suspect; so does
put_pid(p->signal->tty_old_pgrp); /* A noop */
not because of the comment, but because tty_old_pgrp cannot be non-NULL here: 1. The session leader's tty_old_pgrp is only assigned non-NULL if its controlling terminal is hung up. 2. The tty cannot be hung up more than once. 3. If the session leader changes the controlling tty via ioctl(TIOCSCTTY), __proc_set_tty() will put_pid(tty_old_pgrp) and reset it to NULL. [so tty_old_pgrp is NULL on a subsequent hangup of the new controlling tty]. 4. If the session leader drops the controlling terminal association via ioctl(TIOCNOTTY), disassociate_tty() will put_pid(tty_old_pgrp) and reset it to NULL. [Assuming the race mentioned above is fixed, then there is no controlling tty to hangup.]
What about replacing put_pid(p->signal->tty_old_pgrp); /* A noop */ with WARN_ON(p->signal->tty_old_pgrp); ?
And fixing the FIXME in no_tty()?
Regards, Peter Hurley
> --- x/drivers/tty/tty_io.c > +++ x/drivers/tty/tty_io.c > @@ -569,8 +569,7 @@ static int tty_signal_session_leader(str > put_pid(p->signal->tty_old_pgrp); /* A noop */ > spin_lock(&tty->ctrl_lock); > tty_pgrp = get_pid(tty->pgrp); > - if (tty->pgrp) > - p->signal->tty_old_pgrp = get_pid(tty->pgrp); > + p->signal->tty_old_pgrp = get_pid(tty->pgrp); > spin_unlock(&tty->ctrl_lock); > spin_unlock_irq(&p->sighand->siglock); > } while_each_pid_task(tty->session, PIDTYPE_SID, p);
| |