tty_io: fix race in master pty close/slave pty close path

This patch fixes a possible race that leads to double freeing an idr index.
 When the master begin to close, release_dev() is called and then
pty_close() is called:

        if (tty->driver->close)
                tty->driver->close(tty, filp);

This is done without helding any locks other than BKL.  Inside pty_close(),
being a master close, the devpts entry will be removed:

#ifdef CONFIG_UNIX98_PTYS
                if (tty->driver == ptm_driver)
                        devpts_pty_kill(tty->index);
#endif

But devpts_pty_kill() will call get_node() that may sleep while waiting for
&devpts_root->d_inode->i_sem.  When this happens and the slave is being
opened, tty_open() just found the driver and index:

        driver = get_tty_driver(device, &index);
        if (!driver) {
                mutex_unlock(&tty_mutex);
                return -ENODEV;
        }

This part of the code is already protected under tty_mute.  The problem is
that the slave close already got an index.  Then init_dev() is called and
blocks waiting for the same &devpts_root->d_inode->i_sem.

When the master close resumes, it removes the devpts entry, and the
relation between idr index and the tty is gone.  The master then sleeps
waiting for the tty_mutex on release_dev().

Slave open resumes and found no tty for that index.  As result, a NULL tty
is returned and init_dev() doesn't flow to fast_track:

        /* check whether we're reopening an existing tty */
        if (driver->flags & TTY_DRIVER_DEVPTS_MEM) {
                tty = devpts_get_tty(idx);
                if (tty && driver->subtype == PTY_TYPE_MASTER)
                        tty = tty->link;
        } else {
                tty = driver->ttys[idx];
        }
        if (tty) goto fast_track;

The result of this, is that a new tty will be created and init_dev() returns
sucessfull. After returning, tty_mutex is dropped and master close may resume.

Master close finds it's the only use and both sides are closing, then releases
the tty and the index. At this point, the idr index is free, but slave still
has it.

Slave open then calls pty_open() and finds that tty->link->count is 0,
because there's no master and returns error.  Then tty_open() calls
release_dev() which executes without any warning, as it was a case of last
slave close when the master is already closed (master->count == 0,
slave->count == 1).  The tty is then released with the already released idr
index.

This normally would only issue a warning on idr_remove() but in case of a
customer's critical application, it's never too simple:

thread1: opens master, gets index X
thread1: begin closing master
thread2: begin opening slave with index X
thread1: finishes closing master, index X released
thread3: opens master, gets index X, just released
thread2: fails opening slave, releases index X         <----
thread4: opens master, gets index X, init_dev() then find an already in use
         and healthy tty and fails

If no more indexes are released, ptmx_open() will keep failing, as the
first free index available is X, and it will make init_dev() fail because
you're trying to "reopen a master" which isn't valid.

The patch notices when this race happens and make init_dev() fail
imediately.  The init_dev() function is called with tty_mutex held, so it's
safe to continue with tty till the end of function because release_dev()
won't make any further changes without grabbing the tty_mutex.

Without the patch, on some machines it's possible get easily idr warnings
like this one:

idr_remove called for id=15 which is not allocated.
 [<c02555b9>] idr_remove+0x139/0x170
 [<c02a1b62>] release_mem+0x182/0x230
 [<c02a28e7>] release_dev+0x4b7/0x700
 [<c02a0ea7>] tty_ldisc_enable+0x27/0x30
 [<c02a1e64>] init_dev+0x254/0x580
 [<c02a0d64>] check_tty_count+0x14/0xb0
 [<c02a4f05>] tty_open+0x1c5/0x340
 [<c02a4d40>] tty_open+0x0/0x340
 [<c017388f>] chrdev_open+0xaf/0x180
 [<c017c2ac>] open_namei+0x8c/0x760
 [<c01737e0>] chrdev_open+0x0/0x180
 [<c0167bc9>] __dentry_open+0xc9/0x210
 [<c0167e2c>] do_filp_open+0x5c/0x70
 [<c0167a91>] get_unused_fd+0x61/0xd0
 [<c0167e93>] do_sys_open+0x53/0x100
 [<c0167f97>] sys_open+0x27/0x30
 [<c010303b>] syscall_call+0x7/0xb

using this test application available on:
 http://www.ruivo.org/~aris/pty_sodomizer.c

Signed-off-by: Aristeu Sergio Rozanski Filho <aris@ruivo.org>
Signed-off-by: Adrian Bunk <bunk@stusta.de>

diff --git a/drivers/char/tty_io.c b/drivers/char/tty_io.c
index 34d31263d330dda13650df40cb215884e7ae2f01..6fea1c3421bc90c1e47d19cb094ab205a3270c27 100644 (file)
--- a/drivers/char/tty_io.c
+++ b/drivers/char/tty_io.c
@@ -1457,6 +1457,20 @@ static int init_dev(struct tty_driver *driver, int idx,
        /* check whether we're reopening an existing tty */
        if (driver->flags & TTY_DRIVER_DEVPTS_MEM) {
                tty = devpts_get_tty(idx);
+               /*
+                * If we don't have a tty here on a slave open, it's because
+                * the master already started the close process and there's
+                * no relation between devpts file and tty anymore.
+                */
+               if (!tty && driver->subtype == PTY_TYPE_SLAVE) {
+                       retval = -EIO;
+                       goto end_init;
+               }
+               /*
+                * It's safe from now on because init_dev() is called with
+                * tty_mutex held and release_dev() won't change tty->count
+                * or tty->flags without having to grab tty_mutex
+                */
                if (tty && driver->subtype == PTY_TYPE_MASTER)
                        tty = tty->link;
        } else {