ptp: fix the race between the release of ptp_clock and cdev

In a case when a ptp chardev (like /dev/ptp0) is open but an underlying
device is removed, closing this file leads to a race. This reproduces
easily in a kvm virtual machine:

ts# cat openptp0.c
int main() { ... fp = fopen("/dev/ptp0", "r"); ... sleep(10); }
ts# uname -r
5.5.0-rc3-46cf053e
ts# cat /proc/cmdline
... slub_debug=FZP
ts# modprobe ptp_kvm
ts# ./openptp0 &
[1] 670
opened /dev/ptp0, sleeping 10s...
ts# rmmod ptp_kvm
ts# ls /dev/ptp*
ls: cannot access '/dev/ptp*': No such file or directory
ts# ...woken up
[   48.010809] general protection fault: 0000 [#1] SMP
[   48.012502] CPU: 6 PID: 658 Comm: openptp0 Not tainted 5.5.0-rc3-46cf053e #25
[   48.014624] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), ...
[   48.016270] RIP: 0010:module_put.part.0+0x7/0x80
[   48.017939] RSP: 0018:ffffb3850073be00 EFLAGS: 00010202
[   48.018339] RAX: 000000006b6b6b6b RBX: 6b6b6b6b6b6b6b6b RCX: ffff89a476c00ad0
[   48.018936] RDX: fffff65a08d3ea08 RSI: 0000000000000247 RDI: 6b6b6b6b6b6b6b6b
[   48.019470] ...                                              ^^^ a slub poison
[   48.023854] Call Trace:
[   48.024050]  __fput+0x21f/0x240
[   48.024288]  task_work_run+0x79/0x90
[   48.024555]  do_exit+0x2af/0xab0
[   48.024799]  ? vfs_write+0x16a/0x190
[   48.025082]  do_group_exit+0x35/0x90
[   48.025387]  __x64_sys_exit_group+0xf/0x10
[   48.025737]  do_syscall_64+0x3d/0x130
[   48.026056]  entry_SYSCALL_64_after_hwframe+0x44/0xa9
[   48.026479] RIP: 0033:0x7f53b12082f6
[   48.026792] ...
[   48.030945] Modules linked in: ptp i6300esb watchdog [last unloaded: ptp_kvm]
[   48.045001] Fixing recursive fault but reboot is needed!

This happens in:

static void __fput(struct file *file)
{   ...
    if (file->f_op->release)
        file->f_op->release(inode, file); <<< cdev is kfree'd here
    if (unlikely(S_ISCHR(inode->i_mode) && inode->i_cdev != NULL &&
             !(mode & FMODE_PATH))) {
        cdev_put(inode->i_cdev); <<< cdev fields are accessed here

Namely:

__fput()
  posix_clock_release()
    kref_put(&clk->kref, delete_clock) <<< the last reference
      delete_clock()
        delete_ptp_clock()
          kfree(ptp) <<< cdev is embedded in ptp
  cdev_put
    module_put(p->owner) <<< *p is kfree'd, bang!

Here cdev is embedded in posix_clock which is embedded in ptp_clock.
The race happens because ptp_clock's lifetime is controlled by two
refcounts: kref and cdev.kobj in posix_clock. This is wrong.

Make ptp_clock's sysfs device a parent of cdev with cdev_device_add()
created especially for such cases. This way the parent device with its
ptp_clock is not released until all references to the cdev are released.
This adds a requirement that an initialized but not exposed struct
device should be provided to posix_clock_register() by a caller instead
of a simple dev_t.

This approach was adopted from the commit 72139dfa2464 ("watchdog: Fix
the race between the release of watchdog_core_data and cdev"). See
details of the implementation in the commit 233ed09d7fda ("chardev: add
helper function to register char devs with a struct device").

Link: https://lore.kernel.org/linux-fsdevel/20191125125342.6189-1-vdronov@redhat.com/T/#u
Analyzed-by: Stephen Johnston <sjohnsto@redhat.com>
Analyzed-by: Vern Lovejoy <vlovejoy@redhat.com>
Signed-off-by: Vladis Dronov <vdronov@redhat.com>
Acked-by: Richard Cochran <richardcochran@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
4 files changed