Re: Lost events in older kernels

[Henrik Rydberg <rydberg@euromail.se>]

Dmitry Torokhov wrote:
> On Sat, May 22, 2010 at 10:57:42PM +0200, Henrik Rydberg wrote:
>> Dmitry Torokhov wrote:
>>> On Sat, May 22, 2010 at 12:27:15PM +0200, Henrik Rydberg wrote:
>>>> Rafi Rubin wrote:
>>>>> -----BEGIN PGP SIGNED MESSAGE-----
>>>>> Hash: SHA1
>>>>>
>>>>> On 05/22/10 03:42, Dmitry Torokhov wrote:
>>>>>> On Sat, May 22, 2010 at 03:06:07AM -0400, Rafi Rubin wrote:
>>>>>>> I'm playing with a project with a 2.6.29 kernel, and the userspace application
>>>>>>> seems to miss some events.  Is there a particular fix that improved the handling?
>>>>>>>
>>>>>>> Also tried catting the dev to a file while testing, and the dump also is missing
>>>>>>> some events.
>>>>>>>
>>>>>> No "interesting" patches went into evdev for a few release now...
>>>>>>
>>>>>> Hm, could it be that event queue is overflowing before userspace gets a
>>>>>> chance to empty it. What kind of event rate are we talking here?
>>>>>>
>>>>> Quite possibly.  It is a multitouch device and we know Henrik's been concerned
>>>>> with the load for a while.
>>>>>
>>>>> So to put some numbers behind his fears:
>>>>>
>>>>> 146668 hid events processed
>>>>> 24952 evdev events captured with a cat
>>>>> 30 seconds (give or take).
>> Thanks for these numbers, Rafi.
>>
>>>>> This is for a mix of different numbers of fingers, but continuous use for those
>>>>> 30 seconds.  And X was running and reading the dev node too.
>>>> Others have experienced this too. Mika has a patch for this, increasing the
>>>> (kernel) evdev buffer quite a bit, from 64 to 256. I believe the reason it is
>>>> not sent upstream is because it increases the footprint by 3 Kb. Perhaps a
>>>> dynamic solution could work here.
>>>>
>>> Yes, if input devices could hint handlers about their "packet size"
>>> then evdev could size it's event queue accordingly. I'd say we need to
>>> keep about 8 packets worth of data (number is pulled right out of my
>>> behind ;) )...
>>>
>> Yeah. The bcm5974 driver sends 8 events per finger, plus some ST stuff. With
>> four fingers on the pad, that amounts to roughly 40 events per packet.
>>
>> 146668 / 24952 = 5.9
>> 40 * 8 / 64 = 5.0
>>
>> Looks reasonable, doesn't it?
>>
>> Regarding the packet size hinting, the handler already knows which events are
>> potentially being sent, and could produce a reasonable buffer size from it. In
>> particular if it knows which events are bypassing filtering. ;-)
>>
> 
> Yes, but the driver knows for sure. Why making one guess when another
> can tell?
> 

I was thinking that the data stream itself should be enough to dynamically
adjust the buffer size, given an initial appropriate size. In theory, that could
even be better than asking the driver. It would also avoid changing all drivers.
Just a thought. Either way, the problem seems real.

Henrik