Re: [RFC PATCH] usb: hub: Disable autosuspend before disabling usb device

[Guenter Roeck]

On Fri, Mar 17, 2017 at 04:07:14PM -0400, Alan Stern wrote:
> On Fri, 17 Mar 2017, Guenter Roeck wrote:
> 
> > While running a bind/unbind stress test with the dwc3 usb driver on rk3399,
> > the following crash was observed.
> > 
> > Unable to handle kernel NULL pointer dereference at virtual address 00000218
> > pgd = ffffffc00165f000
> > [00000218] *pgd=000000000174f003, *pud=000000000174f003,
> > 				*pmd=0000000001750003, *pte=00e8000001751713
> > Internal error: Oops: 96000005 [#1] PREEMPT SMP
> > Modules linked in: uinput uvcvideo videobuf2_vmalloc cmac
> > ipt_MASQUERADE nf_nat_masquerade_ipv4 iptable_nat nf_nat_ipv4 nf_nat rfcomm
> > xt_mark fuse bridge stp llc zram btusb btrtl btbcm btintel bluetooth
> > ip6table_filter mwifiex_pcie mwifiex cfg80211 cdc_ether usbnet r8152 mii joydev
> > snd_seq_midi snd_seq_midi_event snd_rawmidi snd_seq snd_seq_device ppp_async
> > ppp_generic slhc tun
> > CPU: 1 PID: 29814 Comm: kworker/1:1 Not tainted 4.4.52 #507
> > Hardware name: Google Kevin (DT)
> > Workqueue: pm pm_runtime_work
> > task: ffffffc0ac540000 ti: ffffffc0af4d4000 task.ti: ffffffc0af4d4000
> > PC is at autosuspend_check+0x74/0x174
> > LR is at autosuspend_check+0x70/0x174
> > ...
> > Call trace:
> > [<ffffffc00080dcc0>] autosuspend_check+0x74/0x174
> > [<ffffffc000810500>] usb_runtime_idle+0x20/0x40
> > [<ffffffc000785ae0>] __rpm_callback+0x48/0x7c
> > [<ffffffc000786af0>] rpm_idle+0x1e8/0x498
> > [<ffffffc000787cdc>] pm_runtime_work+0x88/0xcc
> > [<ffffffc000249bb8>] process_one_work+0x390/0x6b8
> > [<ffffffc00024abcc>] worker_thread+0x480/0x610
> > [<ffffffc000251a80>] kthread+0x164/0x178
> > [<ffffffc0002045d0>] ret_from_fork+0x10/0x40
> > 
> > Source:
> > 
> > (gdb) l *0xffffffc00080dcc0
> > 0xffffffc00080dcc0 is in autosuspend_check
> > (drivers/usb/core/driver.c:1778).
> > 1773		/* We don't need to check interfaces that are
> > 1774		 * disabled for runtime PM.  Either they are unbound
> > 1775		 * or else their drivers don't support autosuspend
> > 1776		 * and so they are permanently active.
> > 1777		 */
> > 1778		if (intf->dev.power.disable_depth)
> > 1779			continue;
> > 1780		if (atomic_read(&intf->dev.power.usage_count) > 0)
> > 1781			return -EBUSY;
> > 1782		w |= intf->needs_remote_wakeup;
> > 
> > Code analysis shows that intf is set to NULL in usb_disable_device() prior
> > to setting actconfig to NULL. At the same time, usb_runtime_idle() does not
> > lock the usb device, and neither does any of the functions in the
> > traceback. This means that there is no protection against a race condition
> > where usb_disable_device() is removing dev->actconfig->interface[] pointers
> > while those are being accessed from autosuspend_check() and possibly by
> > other callers.
> 
> That is an interesting race, one that had previously escaped my notice.
> 
> > Explicitly disable autosuspend in usb_disconnect() before calling
> > usb_disable_device(). This doesn't fix the race for good, but it ensures
> > that the pm runtime worker doesn't call usb_runtime_idle() on the interface
> > that is being removed, and thus avoids the race in the affected code path.
> > 
> > Signed-off-by: Guenter Roeck <linux@xxxxxxxxxxxx>
> > ---
> > This is another interesting situation. As mentioned above, the patch doesn't
> > really fix the race problem. On the other side, fixing it for good would
> > (probably) be much more complex. I still see the race after applying this
> > patch, but it happens maybe once a day vs. several times per hour.
> > 
> > Marked as RFC in the hope that someone has an idea for a better fix.
> > I tried clearing udev->actconfig prior to removing the interfaces
> > in usb_disable_device(), but that alone didn't help; it does not
> > resolve the race condition either, and still results in the crash.
> > 
> > The only clean solution I can think of would be to protect accesses
> > to dev->actconfig with a spinlock or mutex, and to make sure that the
> > lock is held during read accesses and that dev->actconfig is cleared
> > before releasing the lock on write accesses. I'll be happy to do that
> > if it is the way to go, but I would like some feedback before I give it
> > a try.
> 
> I think the right thing to do is test udev->state at the start of 
> autosuspend_check(), much like the test at the start of 
> usb_suspend_both().  udev->state gets set to USB_STATE_NOTATTACHED in 
> usb_disconnect() before usb_disable_device() is called.
> 
> Then there will be no need for usb_disable_autosuspend(), although 
> pm_runtime_barrier() will still be necessary.
> 
Makes sense. I'll give it a try.

Thanks,
Guenter