Re: [PATCH] HID: usbhid: do not sleep when opening device

[Dmitry Torokhov]

On Fri, May 29, 2020 at 06:22:49PM -0700, Guenter Roeck wrote:
> On Fri, May 29, 2020 at 6:09 PM Dmitry Torokhov
> <dmitry.torokhov@xxxxxxxxx> wrote:
> >
> > On Fri, May 29, 2020 at 05:48:26PM -0700, Guenter Roeck wrote:
> > > On Fri, May 29, 2020 at 4:50 PM Nicolas Boichat <drinkcat@xxxxxxxxxxxx> wrote:
> > > >
> > > > On Sat, May 30, 2020 at 3:59 AM Dmitry Torokhov
> > > > <dmitry.torokhov@xxxxxxxxx> wrote:
> > > > >
> > > > > usbhid tries to give the device 50 milliseconds to drain its queues
> > > > > when opening the device, but does it naively by simply sleeping in open
> > > > > handler, which slows down device probing (and thus may affect overall
> > > > > boot time).
> > > > >
> > > > > However we do not need to sleep as we can instead mark a point of time
> > > > > in the future when we should start processing the events.
> > > > >
> > > > > Reported-by: Nicolas Boichat <drinkcat@xxxxxxxxxxxx>
> > > > > Signed-off-by: Dmitry Torokhov <dmitry.torokhov@xxxxxxxxx>
> > > > > ---
> > > > >  drivers/hid/usbhid/hid-core.c | 27 +++++++++++++++------------
> > > > >  drivers/hid/usbhid/usbhid.h   |  1 +
> > > > >  2 files changed, 16 insertions(+), 12 deletions(-)
> > > > >
> > > > > diff --git a/drivers/hid/usbhid/hid-core.c b/drivers/hid/usbhid/hid-core.c
> > > > > index c7bc9db5b192..e69992e945b2 100644
> > > > > --- a/drivers/hid/usbhid/hid-core.c
> > > > > +++ b/drivers/hid/usbhid/hid-core.c
> > > > > @@ -95,6 +95,19 @@ static int hid_start_in(struct hid_device *hid)
> > > > >                                 set_bit(HID_NO_BANDWIDTH, &usbhid->iofl);
> > > > >                 } else {
> > > > >                         clear_bit(HID_NO_BANDWIDTH, &usbhid->iofl);
> > > > > +
> > > > > +                       if (test_and_clear_bit(HID_RESUME_RUNNING,
> > > > > +                                              &usbhid->iofl)) {
> > > > > +                               /*
> > > > > +                                * In case events are generated while nobody was
> > > > > +                                * listening, some are released when the device
> > > > > +                                * is re-opened. Wait 50 msec for the queue to
> > > > > +                                * empty before allowing events to go through
> > > > > +                                * hid.
> > > > > +                                */
> > > > > +                               usbhid->input_start_time = jiffies +
> > > > > +                                                          msecs_to_jiffies(50);
> > > > > +                       }
> > > > >                 }
> > > > >         }
> > > > >         spin_unlock_irqrestore(&usbhid->lock, flags);
> > > > > @@ -280,7 +293,8 @@ static void hid_irq_in(struct urb *urb)
> > > > >                 if (!test_bit(HID_OPENED, &usbhid->iofl))
> > > > >                         break;
> > > > >                 usbhid_mark_busy(usbhid);
> > > > > -               if (!test_bit(HID_RESUME_RUNNING, &usbhid->iofl)) {
> > > > > +               if (!test_bit(HID_RESUME_RUNNING, &usbhid->iofl) &&
> > > > > +                   time_after(jiffies, usbhid->input_start_time)) {
> > > >
> > > > Are we worried about jiffies overflowing (32-bit@1000Hz is "only" 49.7 days...)
> > > >
> > >
> > > time_after() is overflow-safe. That is why it is used and jiffies is
> > > not compared directly.
> >
> > Well, it is overflow safe, but still can not measure more than 50 days,
> > so if you have a device open for 50+ days there will be a 50msec gap
> > where it may lose events.
> >
> 
> Or you could explicitly use 64-bit jiffies.

Indeed.

Jiri, Benjamin, do you have preference between jiffies64 and ktime_t? I
guess jiffies64 is a tiny bit less expensive.

Thanks.

-- 
Dmitry