RE: [PATCH] usb: host: ohci: fix a race condition between shutdown and irq

[Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>]

Hi Alan,

> From: Alan Stern, Sent: Saturday, August 24, 2019 12:33 AM
> 
> On Fri, 23 Aug 2019, Yoshihiro Shimoda wrote:
> 
> > This patch fixes an issue that the following error is
> > possible to happen when ohci hardware causes an interruption
> > and the system is shutting down at the same time.
> >
> > [   34.851754] usb 2-1: USB disconnect, device number 2
> > [   35.166658] irq 156: nobody cared (try booting with the "irqpoll" option)
> > [   35.173445] CPU: 0 PID: 22 Comm: kworker/0:1 Not tainted 5.3.0-rc5 #85
> > [   35.179964] Hardware name: Renesas Salvator-X 2nd version board based on r8a77965 (DT)
> > [   35.187886] Workqueue: usb_hub_wq hub_event
> > [   35.192063] Call trace:
> > [   35.194509]  dump_backtrace+0x0/0x150
> > [   35.198165]  show_stack+0x14/0x20
> > [   35.201475]  dump_stack+0xa0/0xc4
> > [   35.204785]  __report_bad_irq+0x34/0xe8
> > [   35.208614]  note_interrupt+0x2cc/0x318
> > [   35.212446]  handle_irq_event_percpu+0x5c/0x88
> > [   35.216883]  handle_irq_event+0x48/0x78
> > [   35.220712]  handle_fasteoi_irq+0xb4/0x188
> > [   35.224802]  generic_handle_irq+0x24/0x38
> > [   35.228804]  __handle_domain_irq+0x5c/0xb0
> > [   35.232893]  gic_handle_irq+0x58/0xa8
> > [   35.236548]  el1_irq+0xb8/0x180
> > [   35.239681]  __do_softirq+0x94/0x23c
> > [   35.243253]  irq_exit+0xd0/0xd8
> > [   35.246387]  __handle_domain_irq+0x60/0xb0
> > [   35.250475]  gic_handle_irq+0x58/0xa8
> > [   35.254130]  el1_irq+0xb8/0x180
> > [   35.257268]  kernfs_find_ns+0x5c/0x120
> > [   35.261010]  kernfs_find_and_get_ns+0x3c/0x60
> > [   35.265361]  sysfs_unmerge_group+0x20/0x68
> > [   35.269454]  dpm_sysfs_remove+0x2c/0x68
> > [   35.273284]  device_del+0x80/0x370
> > [   35.276683]  hid_destroy_device+0x28/0x60
> > [   35.280686]  usbhid_disconnect+0x4c/0x80
> > [   35.284602]  usb_unbind_interface+0x6c/0x268
> > [   35.288867]  device_release_driver_internal+0xe4/0x1b0
> > [   35.293998]  device_release_driver+0x14/0x20
> > [   35.298261]  bus_remove_device+0x110/0x128
> > [   35.302350]  device_del+0x148/0x370
> > [   35.305832]  usb_disable_device+0x8c/0x1d0
> > [   35.309921]  usb_disconnect+0xc8/0x2d0
> > [   35.313663]  hub_event+0x6e0/0x1128
> > [   35.317146]  process_one_work+0x1e0/0x320
> > [   35.321148]  worker_thread+0x40/0x450
> > [   35.324805]  kthread+0x124/0x128
> > [   35.328027]  ret_from_fork+0x10/0x18
> > [   35.331594] handlers:
> > [   35.333862] [<0000000079300c1d>] usb_hcd_irq
> > [   35.338126] [<0000000079300c1d>] usb_hcd_irq
> > [   35.342389] Disabling IRQ #156
> >
> > The ohci_shutdown() should hold the spin lock while disabling
> > the interruption and changing the rh_state flag. Note that
> > io_watchdog_func() also calls the ohci_shutdown() and it
> > already held the spin lock, so that the patch makes a new
> > function as _ohci_shutdown().
> 
> I don't understand this description.  It sounds like the OHCI
> controller generates an interrupt request, and then ohci_shutdown()
> disables the interrupt request before the handler can run.  When the
> handler does run, it sees that no interrupts are enabled and so it
> returns IRQ_NOTMINE, leading to the error shown above.
> 
> How will holding the spinlock fix this problem?

I'm sorry for lacking description. I should have described the following
descriptions instead of that. What do you think?

--
ohci_shutdown() disables all the interrupt and rh_state is set to
OHCI_RH_HALTED. In other hand, ohci_irq() is possible to enable
OHCI_INTR_SF and OHCI_INTR_MIE on ohci_irq(). Note that OHCI_INTR_SF
is possible to be set by start_ed_unlink() which is called:
 ohci_irq()
  -> process_done_list()
   -> takeback_td()
    -> start_ed_unlink()

So, ohci_irq() has the following condition, the issue happens by
&ohci->regs->intrenable = OHCI_INTR_MIE | OHCI_INTR_SF and
ohci->rh_state = OHCI_RH_HALTED:

	/* interrupt for some other device? */
	if (ints == 0 || unlikely(ohci->rh_state == OHCI_RH_HALTED))
		return IRQ_NOTMINE;

To fix the issue, ohci_shutdown() holds the spin lock while disabling
the interruption and changing the rh_state flag to prevent reenable
the OHCI_INTR_MIE unexpectedly. Note that io_watchdog_func() also calls
the ohci_shutdown() and it already held the spin lock, so that the patch
makes a new function as _ohci_shutdown().
--

Best regards,
Yoshihiro Shimoda

> Alan Stern