Message-ID: <20150427035207.GA13074@jelly.redhat.com>
Date:	Mon, 27 Apr 2015 13:52:07 +1000
From:	Peter Hutterer <peter.hutterer@...-t.net>
To:	Henrik Rydberg <rydberg@...math.org>
Cc:	Dmitry Torokhov <dmitry.torokhov@...il.com>,
	Benjamin Tissoires <benjamin.tissoires@...hat.com>,
	Jiri Kosina <jkosina@...e.cz>,
	"linux-input@...r.kernel.org" <linux-input@...r.kernel.org>,
	lkml <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v2] Input - mt: Fix input_mt_get_slot_by_key

On Fri, Apr 24, 2015 at 08:26:39AM +0200, Henrik Rydberg wrote:
> Peter,
> 
> It may be a long time ago now, but we had very vocal discussions regarding the
> MT protocol back then, and I am quite sure all the subtleties are well
> understood. In order to fully appreciate the simplicity of the protocol, one
> only needs to stop misintepreting it. In order to do that, please imagine a
> large piece of papers, a set of brushes, and a set of colors.
> 
> The MT protocol tracks brushes. When lines are drawn, positions are updated.
> When the color on the brush changes, the tracking id changes. When the brush is
> lifted, the tracking id becomes -1, the "no color" color. There is a fixed set
> of brushes. The old test application mtview works precisely this way.
> 
> The add/remove protocol tracks colors. Technichally it tracks contacts, which
> are the combined (color, brush) objects, but in this analogy, colors and
> tracking ids are interchangeable. When lines a drawn, a color is first assigned
> to the color and the brush attributes on the color are updated. The positions on
> the color are subsequently updated. To change color, the brush is deassign from
> the first color and then the brush is assign to the new color. To lift, the
> brush is deassign from the color. This is the abstraction the seems to prevail
> in userland at the moment.
> 
> > can't you just slot in an extra event that contains only the
> > ABS_MT_TRACKING_ID -1 for the needed slots, followed by an SYN_REPORT and
> > whatever BTN_TOOL_* updates are needed? You don't need extra slots here,
> > you're just putting one extra event in the queue before handling the new
> > touches as usual.
> 
> So you want to add a rule saying that before a brush changes color, it first has
> to be cleaned. That may look simple enough, but it misses out on several subtleties.
> 
> 1. It is no longer possible to create beautiful contiuous tracks of varying color.
>
> 2. When a brush moves quickly, it will sometimes restart the track with a new
> color. This happens on all hardware with tracking support when you approach the
> sampling limit. It happens in the kernel tracking as well, for the same reasons.
> Incompatible with 1.
> 
> 3. There is a difference between losing track of a brush and lifting a brush.
> This is one of the situations where the add/remove protocol has to create very
> nonintuitive restrictions and rules to cope. The reason starts with 1.
> 
> Forcing tracks (brushes) into the add/remove protocol creates problems that are
> on a more fundamental level than the subtle issues one may hope to resolve.
> 
> > thing is: I've always assumed that a touch is terminated by a -1 event
> > and this hasn't been a problem until very recently.
> 
> We have talked a lot about the differences, they can hardly have escaped anyone
> deeply involved. 

well, they did. I honestly barely recall any conversations about that
protocol. it's now what, 5 years ago? and since the protocol itself worked
mostly fine I never had to think much about it, short of the various issues
with SYN_DROPPED.

I read through the documentation again after the weekend with fresh eyes and
it IMO can be read both ways. Couple that with seeing tracking IDs of -1 for
a couple of years before this issue came up and here we are at the current
situation. So regardless of what will eventually get merged, I recommend
putting an extra sentence in to explicitly spell out that a tracking ID can
change directly from non-negative to non-negative, and when this is allowed
to happen and what it means to the userspace process. And add an example for
it in the examples section.

> It is true that this is not normally a problem. It only becomes
> important when the sampling rate is too low to resolve all the actions we deem
> important.
> 
> > so anything I've ever
> > touched will be broken if we start switching tracking IDs directly.
> > That
> > includes xorg input, libinput and anything that uses libevdev. sorry.
> 
> This has been in the mainline kernel for the last five years, and obvisouly
> still works well most of the time. I sometimes experience glitches in the
> trackpad usage on my laptop, and it probably stems from this issues. It is
> slightly annoying, but not broken. No reason to panic.

what you're seeing is most likely the effect of libevdev.
libevdev discards the extra tracking ID and the driver will think it's the
same touch. which fixes the various stuck touch issues that we had, along
with crashes, and a couple of memory overflows (specifically in synaptics).

> > if the kernel switches from one tracking ID to another directly,
> > libevdev will actually discard the new tracking ID.
> > http://cgit.freedesktop.org/libevdev/tree/libevdev/libevdev.c#n968
> > (sorry again) aside from the warning, it's safe to switch directly though,
> > there shouldn't be any side-effects. 
> > as for fixing this: I can add something to libevdev to allow it but I'll
> > also need to fix up every caller to handle this sequence then, they all rely
> > on the -1. so some stuff will simply break.
> > plus we still have synaptics up to 1.7.x and evdev up to 2.8.x that are
> > pre-libevdev.
> 
> Perhaps this is worth looking at in conjunction with the problem of handling
> lost touches. I am thinking of suspend/resume issues in particular. If the
> system could handle the distinction between a lift and a lost touch, some logic
> would be less complicated and more correct.

the suspend/resume issues that we had were caused by a SYN_DROPPED event
during resume. that is handled transparently by libevdev now which will drop
touches and restart new touches so that the code to handle SYN_DROPPED and
normal events is virtually the same in the caller.

yes, same underlying problem but different trigger and different handling of
the effect in the xorg drivers, e.g. ignoring any touches already present at
resume time since there are no reliable timestamps on them.

> > for other event processing it's tricky as well. if you go from two
> > touches to two new touches you need to send out a BTN_TOOL_DOUBLETAP 0, then
> > 1. if not, a legacy process missed the event completely (synaptics would
> > suffer from that). likewise, without the BTN_TOUCH going to 0 for one frame
> > you'll get coordinate jumps on the pointer emulation.
> 
> I am not so sure about this - the movement in synaptics checks if there has been
> any identity changes.
> 
> > having the tracking ID go -1 and then to a real one in the same frame makes
> > this even worse, because now even the MT-capable processes need to attach
> > flags to each touch whether it intermediately terminated or not.
> 
> This is simply the result of a poor abstraction to begin with. The state tracked
> through the input subsystem is the slot state.
> 
> > The event
> > ordering is not guaranteed, so we don't know until the SYN_REPORT whether
> > we switched from 2 fingers to 1, or 2 fingers to 2 fingers. or possibly
> > three fingers if BTN_TOOL_TRIPLETAP is set which we won't know until the
> > end. That has to be fixed in every caller. and it changes the evdev protocol
> > from "you have the state at SYN_REPORT" to "you need to keep track of some
> > state changes within the frame". that's no good.
> 
> I think this paragraph is mixing what the kernel is conveying with what
> representation this is mapped to in some userland applications. The state
> transfer of an id change without lifts is correctly transferred.
>
> > so summary: switching directly between IDs is doable but requires userspace
> > fixes everywhere. terminating and restarting a contact within the same frame
> > is going to be nasty, let's not do that please.
> 
> I agree - userland should perhaps use a different abstraction than add/remove.
> Why not use the "brushes" abstraction instead?

it's not as simple as saying "we just use a different abstraction". I
recently traded my time-machine for a surfboard, which did wonders for my
sanity. But the drawback is I can't just go back and make the current code
not happen. I understand that having tracking IDs change directly would be
useful to fix that jumping cursor bug. still means that it requires some
core rewrites in some of the most common users of the evdev protocol.

so the list remains: please don't do -1 and new ID in the same frame.
Alternatively, change tracking ID directly and most current software will
skip over the direct tracking ID change and that's fixable over time in
userspace. Explicitly terminate the touch in the kernel means everything
handles it already.

Cheers,
   Peter

> 
> > best solution: the kernel inserts an additional event to terminate all
> > restarted contacts before starting the new ones as normal in the subsequent
> > frame.
> 
> No, this would just be the first step towards an endless sequence of subtle
> issues trying to shoehorn the protocol into a poor abstraction.
> 
> Henrik
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists