I wasn't sure if this is something to squeeze into 3.8, so don't yell if not. At least Sasha can apply this and re-test against trinity. Changes in v2: - Please review "tty: Don't flush buffer when closing ldisc". This patch replaces the earlier "tty: Don't reschedule buffer work while closing". The text of this commit details why not calling n_tty_flush_buffer() is the correct thing to do, so I won't repeat it here. - Jiri's debug patch "tty: debug buffer work race with tty free" has been included (albeit a slightly different version) Jiri, please sign off (or point out what you'd like changed). - The test jig has been included in the commit message for "tty: Don't flush buffer when closing ldisc" as Alan requested. - Ilya Zykov was added as the Signed-off-by: for the test jig in that same commit message. - Sasha Levin was added as the Reported-by: in that same patch. This patch series addresses the causes of flush_to_ldisc accessing the tty after freeing. The most common cause stems from the n_tty_close() path spuriously scheduling buffer work, when the ldisc has already been halted. This is fixed in 'tty: Don't flush buffer when closing ldisc' The other causes have a central theme: incorrect order-of-operations when halting a line discipline. In general, to prevent buffer work from being scheduled requires: 1. Disallowing further ldisc references 2. Waiting for all existing ldisc references to be released 3. Cancelling existing buffer work If the wait takes place _after_ cancellation, then new work can be scheduled by existing holder(s) of ldisc reference(s). That's bad. Halting the line discipline is performed when, - hanging up the tty (tty_ldisc_hangup()) - TIOCSETD ioctl (tty_set_ldisc()) - finally closing the tty (pair) (tty_ldisc_release()) Concurrent halts are governed by the following requirements: 1. tty_ldisc_release is not concurrent with the other two and so does not need lock or state protection during the ldiscs halt. 2. Accesses through tty->ldisc must be protected by the ldisc_mutex. The wait operation checks the user count (ldisc references) in tty->ldisc->users. 3. tty_set_ldisc() reschedules buffer work that was pending when the ldiscs were halted. That must be an atomic operation in conjunction with re-enabling the ldisc -- which necessitates locking concurrent halts (tty_ldisc_release is exempt here) 4. The legacy mutex cannot be held while waiting for ldisc reference(s) release -or- for cancelling buffer work. 5. Because of #4, the legacy mutex must be dropped prior to or during the halt. Which means reacquiring after the halt. But to preserve lock order the ldisc_mutex must be dropped and reacquired after reacquiring the legacy mutex. 6. Because of #5, the ldisc state may have changed while the ldisc mutex was dropped. Note: this series does not include the lock correction initially reported on by Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx> here https://lkml.org/lkml/2012/11/21/267 . I commented on the latest version here https://lkml.org/lkml/2012/12/3/362 Peter Hurley (11): tty: debug buffer work race with tty free tty: WARN if buffer work racing with tty free tty: Add diagnostic for halted line discipline tty: Refactor n_tty_flush_buffer tty: Don't flush buffer when closing ldisc tty: Refactor wait for ldisc refs out of tty_ldisc_hangup() tty: Remove unnecessary re-test of ldisc ref count tty: Fix ldisc halt sequence on hangup tty: Strengthen no-subsequent-use guarantee of tty_ldisc_halt() tty: Remove unnecessary buffer work flush tty: Halt both ldiscs concurrently drivers/tty/n_tty.c | 34 +++++++--- drivers/tty/pty.c | 3 +- drivers/tty/tty_buffer.c | 5 +- drivers/tty/tty_io.c | 4 +- drivers/tty/tty_ldisc.c | 171 +++++++++++++++++++++++++++++------------------ include/linux/tty.h | 1 + 6 files changed, 139 insertions(+), 79 deletions(-) -- 1.8.0.1 -- To unsubscribe from this list: send the line "unsubscribe linux-serial" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
