A couple of people have hit the WARN_ON() in drivers/char/tty_io.c,
tty_open() that is unhappy about seeing the tty line discipline go away
during the tty hangup. See for example
http://bugzilla.kernel.org/show_bug.cgi?id=14255
and the reason is that we do the tty_ldisc_halt() outside the
ldisc_mutex in order to be able to flush the scheduled work without a
deadlock with vhangup_work.
However, it turns out that we can solve this particular case by
- using "cancel_delayed_work_sync()" in tty_ldisc_halt(), which waits
for just the particular work, rather than synchronizing with any
random outstanding pending work.
This won't deadlock, since the buf.work we synchronize with doesn't
care about the ldisc_mutex, it just flushes the tty ldisc buffers.
- realize that for this particular case, we don't need to wait for any
hangup work, because we are inside the hangup codepaths ourselves.
so as a result we can just drop the flush_scheduled_work() entirely, and
then move the tty_ldisc_halt() call to inside the mutex. That way we
never expose the partially torn down ldisc state to tty_open(), and hold
the ldisc_mutex over the whole sequence.
Reported-by: Ingo Molnar <mingo@elte.hu>
Reported-by: Heinz Diehl <htd@fancy-poultry.org>
Cc: stable@kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Now that the /proc/tty/ldiscs handling doesn't play games with 'struct
ldisc' any more, the only remaining user of 'tty_ldisc_try_get()' is
'tty_ldisc_get()' (note the lack of 'try').
And we're actually much better off folding the logic directly into that
file, since the 'try' part was always about trying to get the ldisc
operations, not the ldisc itself: and making that explicit inside of
'tty_ldisc_get()' clarifies the whole semantics.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>,
Tested-by: Sergey Senozhatsky <sergey.senozhatsky@mail.by>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
The /proc/tty/ldiscs file is totally and utterly un-interested in the
"struct tty_ldisc" structures, and only cares about the underlying ldisc
operations.
So don't make it create a dummy 'struct ldisc' only to get a pointer to
the operations, and then destroy it. Instead, we split up the function
'tty_ldisc_try_get()', and create a 'get_ldops()' helper that just looks
up the ldisc operations based on the ldisc number.
That makes the code simpler to read (smaller and more well-defined
helper functions), and allows the /proc functions to avoid creating that
useless dummy only to immediately free it again.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Tested-by: Sergey Senozhatsky <sergey.senozhatsky@mail.by>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
When I rewrote tty ldisc code to use proper reference counts (commits
65b770468e and cbe9352fa0) in order to avoid a race with hangup, the
test-program that Eric Biederman used to trigger the original problem
seems to have exposed another long-standing bug: the hangup code did the
'tty_ldisc_halt()' to stop any buffer flushing activity, but unlike the
other call sites it never actually flushed any pending work.
As a result, if you get just the right timing, the pending work may be
just about to execute (ie the timer has already triggered and thus
cancel_delayed_work() was a no-op), when we then re-initialize the ldisc
from under it.
That, in turn, results in various random problems, usually seen as a
NULL pointer dereference in run_timer_softirq() or a BUG() in
worker_thread (but it can be almost anything).
Fix it by adding the required 'flush_scheduled_work()' after doing the
tty_ldisc_halt() (this also requires us to move the ldisc halt to before
taking the ldisc mutex in order to avoid a deadlock with the workqueue
executing do_tty_hangup, which requires the mutex).
The locking should be cleaned up one day (the requirement to do this
outside the ldisc_mutex is very annoying, and weakens the lock), but
that's a larger and separate undertaking.
Reported-by: Eric W. Biederman <ebiederm@xmission.com>
Tested-by: Xiaotian Feng <xtfeng@gmail.com>
Tested-by: Yanmin Zhang <yanmin_zhang@linux.intel.com>
Tested-by: Dave Young <hidave.darkstar@gmail.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Greg Kroah-Hartman <gregkh@suse.de>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Use 'atomic_dec_and_lock()' to make sure that we always hold the
tty_ldisc_lock when the ldisc count goes to zero. That way we can never
race against 'tty_ldisc_try()' increasing the count again.
Reported-by: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Tested-by: Sergey Senozhatsky <sergey.senozhatsky@mail.by>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
By using the user count for the actual lifetime rules, we can get rid of
the silly "wait_for_idle" logic, because any busy ldisc will
automatically stay around until the last user releases it. This avoids
a host of odd issues, and simplifies the code.
So now, when the last ldisc reference is dropped, we just release the
ldisc operations struct reference, and free the ldisc.
It looks obvious enough, and it does work for me, but the counting
_could_ be off. It probably isn't (bad counting in the new version would
generally imply that the old code did something really bad, like free an
ldisc with a non-zero count), but it does need some testing, and
preferably somebody looking at it.
With this change, both 'tty_ldisc_put()' and 'tty_ldisc_deref()' are
just aliases for the new ref-counting 'put_ldisc()'. Both of them
decrement the ldisc user count and free it if it goes down to zero.
They're identical functions, in other words.
But the reason they still exist as sepate functions is that one of them
was exported (tty_ldisc_deref) and had a stupid name (so I don't want to
use it as the main name), and the other one was used in multiple places
(and I didn't want to make the patch larger just to rename the users).
In addition to the refcounting, I did do some minimal cleanup. For
example, now "tty_ldisc_try()" actually returns the ldisc it got under
the lock, rather than returning true/false and then the caller would
look up the ldisc again (now without the protection of the lock).
That said, there's tons of dubious use of 'tty->ldisc' without obviously
proper locking or refcounting left. I expressly did _not_ want to try to
fix it all, keeping the patch minimal. There may or may not be bugs in
that kind of code, but they wouldn't be _new_ bugs.
That said, even if the bugs aren't new, the timing and lifetime will
change. For example, some silly code may depend on the 'tty->ldisc'
pointer not changing because they hold a refcount on the 'ldisc'. And
that's no longer true - if you hold a ref on the ldisc, the 'ldisc'
itself is safe, but tty->ldisc may change.
So the proper locking (remains) to hold tty->ldisc_mutex if you expect
tty->ldisc to be stable. That's not really a _new_ rule, but it's an
example of something that the old code might have unintentionally
depended on and hidden bugs.
Whatever. The patch _looks_ sensible to me. The only users of
ldisc->users are:
- get_ldisc() - atomically increment the count
- put_ldisc() - atomically decrements the count and releases if zero
- tty_ldisc_try_get() - creates the ldisc, and sets the count to 1.
The ldisc should then either be released, or be attached to a tty.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Tested-by: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
Tested-by: Sergey Senozhatsky <sergey.senozhatsky@mail.by>
Acked-by: Alan Cox <alan@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
This is pure preparation of changing the ldisc reference counting to be
a true refcount that defines the lifetime of the ldisc. But this is a
purely syntactic change for now to make the next steps easier.
This patch should make no semantic changes at all. But I wanted to make
the ldisc refcount be an atomic (I will be touching it without locks
soon enough), and I wanted to rename it so that there isn't quite as
much confusion between 'ldo->refcount' (ldisk operations refcount) and
'ld->refcount' (ldisc refcount itself) in the same file.
So it's now an atomic 'ld->users' count. It still starts at zero,
despite having a reference from 'tty->ldisc', but that will change once
we turn it into a _real_ refcount.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Tested-by: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
Tested-by: Sergey Senozhatsky <sergey.senozhatsky@mail.by>
Acked-by: Alan Cox <alan@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
We can get a situation where a hangup occurs during or after a close. In
that case the ldisc gets disposed of by the close and the hangup then
explodes.
Signed-off-by: Alan Cox <alan@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* Remove smp_lock.h from files which don't need it (including some headers!)
* Add smp_lock.h to files which do need it
* Make smp_lock.h include conditional in hardirq.h
It's needed only for one kernel_locked() usage which is under CONFIG_PREEMPT
This will make hardirq.h inclusion cheaper for every PREEMPT=n config
(which includes allmodconfig/allyesconfig, BTW)
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently we reinit the ldisc on final tty close which is what the old code
did to ensure that if the device retained its termios settings then it had the
right ldisc. tty_ldisc_reinit does that but also leaves us with the reset
ldisc reference which is then leaked.
At this point we know the port will be recycled so we can kill the ldisc
off completely rather than try and add another ldisc free up when the kref
count hits zero.
At this point it is safe to keep the ldisc closed as tty_ldisc waiting
methods are only used from the user side, and as the final close we are
the last such reference. Interrupt/driver side methods will always use the
non wait version and get back a NULL.
Found with kmemleak and investigated/identified by Catalin Marinas.
Signed-off-by: Alan Cox <alan@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We have a tty_ldisc file now so put tty_ldisc_flush in the right place
Signed-off-by: Alan Cox <alan@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There are several pretty much unfixable races in the old ldisc code, especially
with respect to pty behaviour and also to hangup. It's easier to rewrite the
code than simply try and patch it up.
This patch
- splits the ldisc from the tty (so we will be able to refcount it more cleanly
later)
- introduces a mutex lock for ldisc changing on an active device
- fixes the complete mess that hangup caused
- implements hopefully correct setldisc/close/hangup locking
There are still some problems around pty pairs that have always been there but
at least it is now possible to understand the code and fix further problems.
This fixes the following known bugs
- hang up can leak ldisc references
- hang up may not call open/close on ldisc in a matched way
- pty/tty pairs can deadlock during an ldisc change
- reading the ldisc proc files can cause every ldisc to be loaded
and probably a few other of the mysterious ldisc race reports.
I'm sure it also adds the odd new one.
Signed-off-by: Alan Cox <alan@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Before trying to tackle the ldisc bugs the code needs to be a good deal
more readable, so do the simple extractions of routines first.
Signed-off-by: Alan Cox <alan@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently function tty_ldisc_get() tries to load an ldisc driver module
only when tty_ldisc_try_get() returns -EAGAIN. This happens only if
module is being unloaded. If ldisc module is not loaded
tty_ldisc_try_get() returns -EINVAL and this case is not handled in
tty_ldisc_get(), so request_module() is not called.
Attached patch fixes this by calling request_module() if
tty_ldisc_try_get() returned any error code.
I discovered this when my UMTS modem stopped working with 2.6.27-rc1
because module ppp_async was not loaded.
Signed-off-by: Eugeniy Meshcheryakov <eugen@debian.org>
Signed-off-by: Alan Cox <alan@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
