Hi Sakari,
On Sunday 10 Jul 2016 01:03:09 Sakari Ailus wrote:
> On Sat, Jul 09, 2016 at 10:29:03PM +0300, Laurent Pinchart wrote:
> > On Monday 09 May 2016 16:16:26 Sakari Ailus wrote:
> >> Laurent Pinchart wrote:
> >>> On Wednesday 04 May 2016 16:09:51 Sakari Ailus wrote:
> >>>> Refactor copying the IOCTL argument structs from the user space and
> >>>> back, in order to reduce code copied around and make the
> >>>> implementation more robust.
> >>>>
> >>>> As a result, the copying is done while not holding the graph mutex.
> >>>>
> >>>> Signed-off-by: Sakari Ailus <sakari.ailus@xxxxxxxxxxxxxxx>
> >>>> ---
> >>>> since v2:
> >>>>
> >>>> - Remove function to calculate maximum argument size, replace by a
> >>>> char array of 256 or kmalloc() if that's too small.
> >>>>
> >>>> drivers/media/media-device.c | 194 ++++++++++++++-------------------
> >>>> 1 file changed, 94 insertions(+), 100 deletions(-)
> >>>>
> >>>> diff --git a/drivers/media/media-device.c
> >>>> b/drivers/media/media-device.c
> >>>> index 9b5a88d..0797e4b 100644
> >>>> --- a/drivers/media/media-device.c
> >>>> +++ b/drivers/media/media-device.c
> >
> > [snip]
> >
> >>>> @@ -453,10 +432,24 @@ static long __media_device_ioctl(
> >>>>
> >>>> info = &info_array[_IOC_NR(cmd)];
> >>>>
> >>>> + if (_IOC_SIZE(info->cmd) > sizeof(__karg)) {
> >>>> + karg = kmalloc(_IOC_SIZE(info->cmd), GFP_KERNEL);
> >>>> + if (!karg)
> >>>> + return -ENOMEM;
> >>>> + }
> >>>> +
> >>>> + info->arg_from_user(karg, arg, cmd);
> >>>> +
> >>>> mutex_lock(&dev->graph_mutex);
> >>>> - ret = info->fn(dev, arg);
> >>>> + ret = info->fn(dev, karg);
> >>>> mutex_unlock(&dev->graph_mutex);
> >>>>
> >>>> + if (!ret)
> >>>
> >>> How about if (!ret && info->arg_to_user) instead, and getting rid of
> >>> copy_arg_to_user_nop() ?
> >>
> >> I thought of that, but I decided to optimise the common case --- which
> >> is that the argument is copied back and forth. Not copying the argument
> >> back is a very special case, we use it for a single compat IOCTL.
> >>
> >> That said, we could use it for the proper ENUM_LINKS as well. Still that
> >> does not change what's normal.
> >
> > We're talking about one comparison and one branching instruction (that
> > will not be taken in the common case). Is that micro-optimization really
> > worth it in an ioctl path that is not that performance-critical ? If you
> > think it is, could you analyse what the impact of the
> > copy_arg_to_user_nop() function on cache locality is for the common case ?
> > ;-)
>
> I sense a certain amount of insistence in your arguments. Fine, I'll change
> it.
Thanks. I'll change that in the next version of the request API patches I will
send out.
> You might want to send a patch removing video_device_release_empty() as
> well. :-)
Actually we should, but for an entirely different reason : most drivers that
use video_device_release_empty() do so because they believe devm_kzalloc() is
the best invention since sliced bread, but in reality they will crash at
unbind time if userspace holds a reference to the video node.
--
Regards,
Laurent Pinchart
--
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
