Re: [PATCH v5] staging: r8188eu: Remove _enter/_exit_critical_mutex()

[Pavel Skripkin]

On 9/3/21 18:09, Pavel Skripkin wrote:
> On 9/2/21 12:32, Greg Kroah-Hartman wrote:
> > On Sat, Aug 28, 2021 at 01:36:56PM +0200, Fabio M. De Francesco wrote:
> > > Remove _enter_critical_mutex() and _exit_critical_mutex(). They are
> > > unnecessary wrappers, respectively to mutex_lock_interruptible() and
> > > to mutex_unlock(). They also have an odd interface that takes an unused
> > > argument named pirqL of type unsigned long.
> > > The original code enters the critical section if the mutex API is
> > > interrupted while waiting to acquire the lock; therefore it could lead
> > > to a race condition. Use mutex_lock() because it is uninterruptible and
> > > so avoid that above-mentioned potential race condition.
> > >
> > > Tested-by: Pavel Skripkin <paskripkin@xxxxxxxxx>
> > > Reviewed-by: Pavel Skripkin <paskripkin@xxxxxxxxx>
> > > Signed-off-by: Fabio M. De Francesco <fmdefrancesco@xxxxxxxxx>
> > > ---
> > >
> > > v5: Fix a typo in the subject line. Reported by Aakash Hemadri.
> > >
> > > v4: Tested and reviewed by Pavel Skripkin. No changes to the code.
> > >
> > > v3: Assume that the original authors don't expect that
> > > mutex_lock_interruptible() can be really interrupted and then lead to 
> > > a potential race condition. Furthermore, Greg Kroah-Hartman makes me
> > > notice that "[] one almost never needs interruptable locks in a driver".
> > > Therefore, replace the calls to mutex_lock_interruptible() with calls to
> > > mutex_lock() since the latter is uninterruptible and avoid race
> > > conditions without the necessity to handle -EINTR errors.
> >
> > Based on a recent conversation on the linux-usb mailing list, perhaps I
> > was wrong:
> > 	https://lore.kernel.org/r/20210829015825.GA297712@xxxxxxxxxxxxxxxxxxx
> >
> > Can you check what happens with your change when you disconnect the
> > device and these code paths are being called?  That is when you do want
> > the lock interrupted.
> >
> > Yes, the logic still seems wrong, but I don't want to see the code now
> > just lock up entirely with this change as it is a change in how things
> > work from today.
>
> Hi, Greg!
>
> I've retested this patch with lockdep enabled and I actually hit a
> deadlock. It's really my fault to forgot about lockdep while testing v4,
> I am sorry about the situation.
>
> Actually, the disconnect here is not the problem, the problem was in
> original code. Changing mutex_lock_interruptible to mutex_lock just
> helped to discover it.
>
>
> The log:
>
> [  252.063305] WARNING: possible recursive locking detected
> [  252.063642] 5.14.0+ #9 Tainted: G         C
> [  252.063946] --------------------------------------------
> [  252.064282] ip/335 is trying to acquire lock:
> [  252.064560] ffff888009ebad28 (pmutex){+.+.}-{4:4}, at:
> usbctrl_vendorreq+0xc5/0x4a0 [r8188eu]
> [  252.065168]
> [  252.065168] but task is already holding lock:
> [  252.065536] ffffffffc021b3b8 (pmutex){+.+.}-{4:4}, at:
> netdev_open+0x3a/0x5f [r8188eu]
> [  252.066085]
> [  252.066085] other info that might help us debug this:
> [  252.066494]  Possible unsafe locking scenario:
> [  252.066494]
> [  252.066866]        CPU0
> [  252.067025]        ----
> [  252.067184]   lock(pmutex);
> [  252.067367]   lock(pmutex);
> [  252.067548]
> [  252.067548]  *** DEADLOCK ***
> [  252.067548]
> [  252.067920]  May be due to missing lock nesting notation
> [  252.067920]
> [  252.068346] 2 locks held by ip/335:
> [  252.068570]  #0: ffffffffbda94628 (rtnl_mutex){+.+.}-{4:4}, at:
> rtnetlink_rcv_msg+0x1e0/0x660
> [  252.069115]  #1: ffffffffc021b3b8 (pmutex){+.+.}-{4:4}, at:
> netdev_open+0x3a/0x5f [r8188eu]
> [  252.069690]
> [  252.069690] stack backtrace:
> [  252.069968] CPU: 1 PID: 335 Comm: ip Tainted: G         C
> 5.14.0+ #9
> [  252.071111] Call Trace:
> [  252.071273]  dump_stack_lvl+0x45/0x59
> [  252.071513]  __lock_acquire.cold+0x1fe/0x31b
> [  252.072709]  lock_acquire+0x157/0x3c0
> [  252.074445]  __mutex_lock+0xf6/0xc90
> [  252.076294]  usbctrl_vendorreq+0xc5/0x4a0 [r8188eu]
> [  252.076651]  usb_read8+0x68/0x8f [r8188eu]
> [  252.076962]  ? usb_read16+0x8e/0x8e [r8188eu]
> [  252.077287]  _rtw_read8+0x2d/0x32 [r8188eu]
> [  252.077601]  HalPwrSeqCmdParsing+0x143/0x1de [r8188eu]
> [  252.077979]  rtl8188eu_InitPowerOn+0x5a/0xe0 [r8188eu]
> [  252.078352]  rtl8188eu_hal_init+0xe7/0x1008 [r8188eu]
> [  252.078989]  rtw_hal_init+0x38/0xb5 [r8188eu]
> [  252.079317]  _netdev_open+0x282/0x4db [r8188eu]
> [  252.079653]  netdev_open+0x42/0x5f [r8188eu]

Ok, sorry for noise. It's 100% false positive. Why?

There is no pmutex in this driver. But! *All* mutexes are initialied via 
private _rtw_mutex_init() API, which has a struct mutex *pmutex argument.

So, driver registers all mutexes with the same name to lockdep map. Of 
course, lockdep will complain about _any_ nested locking...


I will prepare a patch to fix this *completely wrong* approach...




With regards,
Pavel Skripkin