Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Thu, 1 Feb 2018 10:33:33 -0500 (EST)
From:   Alan Stern <stern@rowland.harvard.edu>
To:     "Yoshida, Shigeru" <Shigeru.Yoshida@windriver.com>
cc:     "Bai, Haiqing" <Haiqing.Bai@windriver.com>,
        "gregkh@linuxfoundation.org" <gregkh@linuxfoundation.org>,
        "linux-usb@vger.kernel.org" <linux-usb@vger.kernel.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH] ohci-hcd: Fix race condition caused by ohci_urb_enqueue()
 and io_watchdog_func()
In-Reply-To: <20180201.144911.2171407625301691891.shigeru.yoshida@windriver.com>
Message-ID: <Pine.LNX.4.44L0.1802011029160.1677-100000@iolanthe.rowland.org>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, 1 Feb 2018, Yoshida, Shigeru wrote:

> Hi Alan,
> 
> Thank you for your commenting.
> 
> On Wed, 31 Jan 2018 11:02:47 -0500, Alan Stern wrote:
> >> To address above scenario, this patch introduces timer_running flag to
> >> ohci_hcd structure.  Setting true to ohci->timer_running indicates
> >> io_watchdog_func() is scheduled or is running.  ohci_urb_enqueue()
> >> checks the flag when it schedules the watchdog (step 4 and 12 above),
> >> so ohci->prev_frame_no is not overwritten while io_watchdog_func() is
> >> running.
> > 
> > Instead of adding an extra flag variable, which has to be kept in sync 
> > with the timer routine, how about defining a special sentinel value for 
> > prev_frame_no?  For example:
> > 
> > #define IO_WATCHDOG_OFF		0xffffff00
> > 
> > Then whenever the timer isn't scheduled or running, set
> > ohci->prev_frame_no to IO_WATCHDOG_OFF.  And instead of testing
> > timer_pending(), compare prev_frame_no to this special value.
> > 
> > I think that approach will be slightly more robust.
> 
> It's reasonable since ohci->prev_frame_no is not used while the
> watchdog timer is stopped.
> 
> I think we must choose an invalid frame number for the special
> sentinel value, but I'm not sure which value is adequate for it.
> Is 0xffffff00 an invalid frame number, otherwise how about simply
> -1(0xffffffff)?

Well, the frame_no register is 32 bits wide, but only the 16 low-order 
bits are meaningful.  ohci_frame_no() strips off the high-order 16 
bits, so any value with one of those bits set would be acceptable.  
(Besides, valid frame numbers only go up to 2047.)

I chose 0xffffff00 because PCI reads from a non-working device 
generally get a value with all the bits set.  But since the upper 16 
bits are masked away anyhow, it doesn't matter.  -1u would be fine.

Alan Stern