Re: [PATCH] [v8]Input: evdev: fix bug of dropping valid packet after syn_dropped event

From: Aniroop Mathur
Date: Sat Jan 23 2016 - 15:05:22 EST


On Sun, Jan 24, 2016 at 12:09 AM, Dmitry Torokhov
<dmitry.torokhov@xxxxxxxxx> wrote:
> On Sat, Jan 23, 2016 at 11:29:29PM +0530, Aniroop Mathur wrote:
>> Hi Mr. Torokhov,
>>
>> On Fri, Jan 22, 2016 at 12:47 AM, Dmitry Torokhov
>> <dmitry.torokhov@xxxxxxxxx> wrote:
>> > Hi Anoroop,
>> >
>> > On Thu, Jan 21, 2016 at 11:07:19PM +0530, Aniroop Mathur wrote:
>> >> If last event dropped in the old queue was EVi_SYN/SYN_REPORT, then lets
>> >> generate EV_SYN/SYN_REPORT immediately after queing EV_SYN/SYN_DROPPED
>> >> so that clients would not ignore next valid full packet events.
>> >>
>> >> Signed-off-by: Aniroop Mathur <a.mathur@xxxxxxxxxxx>
>> >> ---
>> >> drivers/input/evdev.c | 46 ++++++++++++++++++++++++++++++++++------------
>> >> 1 file changed, 34 insertions(+), 12 deletions(-)
>> >>
>> >> diff --git a/drivers/input/evdev.c b/drivers/input/evdev.c
>> >> index e9ae3d5..821b68a 100644
>> >> --- a/drivers/input/evdev.c
>> >> +++ b/drivers/input/evdev.c
>> >> @@ -156,7 +156,12 @@ static void __evdev_flush_queue(struct evdev_client *client, unsigned int type)
>> >> static void __evdev_queue_syn_dropped(struct evdev_client *client)
>> >> {
>> >> struct input_event ev;
>> >> + struct input_event *last_ev;
>> >> ktime_t time;
>> >> + unsigned int mask = client->bufsize - 1;
>> >> +
>> >> + /* capture last event stored in the buffer */
>> >> + last_ev = &client->buffer[(client->head - 1) & mask];
>> >
>> > I have still the same comment. How do you know that event at last_ev
>> > position is in fact valid and unconsumed yet event. Also, you need to
>> > figure out not only if queue contains last SYN event, but also to handle
>> > the case where the queue is empty and client has consumed either full or
>> > partial packet at the time you are queueing the drop.
>> >
>>
>> Could you please explain what you mean exactly so that I could answer it
>> properly?
>>
>> From what I understood, it seems to me that there is no problem related to
>> validity, unconsumption, empty queue, full/partial packet.
>> I would like to explain for clock change request case so that you can know
>> my understanding for your question.
>>
>> Clock change request case:
>>
>> 1.1 About last event being valid and unconsumed:
>> We flush buffer and queue syn_dropped only when buffer is not empty. So there
>> will be always be atleast one event in buffer that is not consumed and is
>> ofcourse valid.
>>
>> 1.2 About queue is empty
>> If not empty, we do not flush or add syn_dropped at all.
>
> Clock type change is not the only time we queue SYN_DROP, the other time
> is when we fail to handle EVIOCG[type] (during which we remove some
> events from the queue). Queue may be empty when these ioctls are issued.
>

yeah, ideally it should be changed to:
ret = bits_to_user(mem, maxbit, maxlen, p, compat);
if (ret < 0)
+ if (client->head != client->tail)
- evdev_queue_syn_dropped(client);
+ evdev_queue_syn_dropped(client);

Firstly, there is a need to follow protocol of SYN_REPORT as mentioned in
next section.

Unless syn_report does not denote end of a packet,
it is okay without this change too because last event stored would be
syn_report and after flushing, syn_report will still be at the end and
with empty queue too if syn_dropped is queued then we have added syn_report
to not ignore upcoming valid packet.

>>
>> 1.3 About consumption of full or partial packet
>> If client has consumed full packet, then buffer will look like,
>> ... X Y S(consumed) ... X Y S
>> As we always store packets keeping buffer lock and not single events, so there
>> will always be syn_report in the end.
>
> We try to pass full packets to clients, but there is no guarantee. We
> only estimate number of events in device's packet, not guarantee that it
> is correct size.
>

As per documentation, SYN_REPORT should be used to separate packets.
* SYN_REPORT:
- Used to synchronize and separate events into packets of input data changes
occurring at the same moment in time. For example, motion of a mouse may set
the REL_X and REL_Y values for one motion, then emit a SYN_REPORT. The next
motion will emit more REL_X and REL_Y values and send another SYN_REPORT.

So I think, there is a need of below change:
file: input.c
function: input_handle_event:
} else if (dev->num_vals >= dev->max_vals - 2) {
- dev->vals[dev->num_vals++] = input_value_sync;
input_pass_values(dev, dev->vals, dev->num_vals);

When the packet is really completed
We have sufficient space in buffer to store more than 1 packet even when
actual packet size is more than max_vals so there seems no need to add
syn_report event here by self. So whenever driver sends syn_report, then only
it will be considered as end of a packet and on exceding max_vals, we can
simply pass to handlers to store those partial events in buffer. And anyways
unless the packet is really completed then only client will send it to
application.

Without following this protocol, we would not be able to find the end of a
packet because if max_vals comes out to 3 but actual packet size is 15, then
in the buffer, there will be many syn_reports within a single packet. So with
both current code and with patch code, there will be trouble in handling
syn_dropped because after one syn_report comes, client will stop ignoring the
events.

>> If syn_dropped is queued here, then queing syn_report is fine.
>> If client has consumed partial packet, then buffer will look like,
>> ... X(consumed) Y S ... S
>> If syn_dropped is queued here, then it is fine to queue syn_report because
>> now new X Y will be reported by driver, and so client will consume new X and Y
>> and that old X will be replaced with new X and this new packet will be sent by
>> client to application. Obviously, client never sends partial events to
>> application and send data packet by packet.
>> So I do not see any trouble here related to partial or full packet.
>>
>> > Also please enumerate what changes you done between version n and n+1 so
>> > I do not have to compare them line by line trying to figure it out
>> > myself.
>> >
>>
>> Difference from v5:
>> Made a mistake about head index in v5.
>> Corrected in v8.
>>
>> evdev_set_clk_type:
>> - client->packet_head = client->head = client->tail;
>> + client->packet_head = client->tail = client->head;
>
> Why does this matter?
>

We must not change the head index. So we need to make packet_head and tail
index same as head index and not same as tail index. After this setting,
we call evdev_queue_syn_dropped where we capture last event as the event
present at (head - 1) index;

Thanks,
Aniroop Mathur

> Thanks.
>
> --
> Dmitry