Message-ID: <6792470.H7eb3IQorJ@wuerfel>
Date:	Tue, 01 Dec 2015 11:47:52 +0100
From:	Arnd Bergmann <arnd@...db.de>
To:	Pingbo Wen <pingbo.wen@...aro.org>
Cc:	y2038@...ts.linaro.org,
	Dmitry Torokhov <dmitry.torokhov@...il.com>,
	aksgarg1989@...il.com, linux-input@...r.kernel.org,
	linux-kernel@...r.kernel.org, linux-api@...r.kernel.org
Subject: Re: [PATCH 0/3] introduce new evdev interface type

On Tuesday 01 December 2015 16:34:00 Pingbo Wen wrote:
> Hi, Arnd
> 
> >>>> 
> >>>> The patch series add three evdev interface type:
> >>>> 
> >>>> - EV_IF_LEGACY
> >>>> send event by input_event. This is the default option, keep kernel
> >>>> backward compatible.
> >>> 
> >>> The problem I see with this approach is that it still breaks any
> >>> legacy source code that is compiled with a new libc that uses 64-bit
> >>> time_t. If we are requiring source code changes for building users
> >>> of input devices with a new libc, we can easily get them to handle
> >>> the overflow (they normally only care about the microsecond portion
> >>> anyway, so it doesn't matter in most cases), or to use monotonic time.
> >> 
> >> I don’t think so.
> >> 
> >> Actually, from the view of userspace, EV_IF_LEGACY interface is work
> >> exactly the same as old evdev. We didn’t change anything in input_event
> >> and input_event_compat. And the problem you said will still be there,
> >> even without those patches.
> > 
> > I think we're still talking past one another. I thought we had established
> > that
> > 
> > 1. the current interface is broken when time_t changes in user space
> 
> What kinds of changes in time_t? Extending it to 64-bits in both kernel
> and userspace? Ok, I get confused here, if there are some sample codes
> or use-cases here, it will help me a lot.

We don't change time_t in the kernel, we just try to replace it
with time64_t, or ktime_t where appropriate.

What I meant is the problem when glibc defines their time_t to
be 'long long' so that user space can be built that runs after
2038. This changes the timeval and timespec definitions, so
a process that tries to use 'struct input_event' has a different
layout compared to what the kernel has.

I though that we had already discussed that multiple times.

> > 2. we can fix it by redefining struct input_event in a way that
> >   is independent of time_t
> > 3. once both user space and kernel are using the same ABI independent
> >   of time_t, we can look improving the timestamps so they don't
> >   overflow
> > 4. the monotonic timestamp interface already avoids the overflow, so
> >   it would be sufficient as a solution for 3.
> > 
> > Where did I lose you here? Did you find any other facts that I
> > was missing? I don't know whether the two new event structures make
> > the interface better in some other way, but it seems mostly unrelated
> > to either of the two problems we already have with time_t (the
> > ABI mismatch, and the use of non-monotonic timestamps).
> 
> It seems we are mismatch here.
> 
> Actually input_composite_event has the similar structure with input_event,
> but with a nicer definition, which can take both monotonic and non-monotonic
> timestamps safely.
> 
> What I assumed here, is that leaving EV_IF_LEGACY as a untouched, deprecated
> interface. If userspace try to adapt to new libc and kernel, it should move
> to new interface. The userspace can do a lazy update, keep the code untouched,
> but suffer the y2038 problem by itself.

Forcing the move to the new API is very ugly, because you can't do it
in a way that works on old kernels as well, unless you then try to support
both APIs at runtime.

Just requiring user space to switch to monotonic time is easily done,
as it can likely work either way.

> We can force kernel using monotonic time in EV_IF_LEGACY interface, and
> making input_event independent from time_t(after evdev has converted to
> input_value, it’s easy to do that), but that also imply userspace
> must change their code to fit this change. If changing userspace code is
> a mandatory option, why not to force them do a complete conversion?

Most user space programs won't care, as they don't even look at the tv_sec
portion, and the goal is to avoid having to change them.

There is still an open question to how exactly we want to get user space
to change.

We could do some compile-time trick by having a union in struct input_event
and mark the existing timeval part as deprecated, so we flag any use of the
32-bit tv_sec member, such as:

struct input_event {
#if !defined(__KERNEL__) && __TIME_T_BITS == __BITS_PER_LONG
        struct timeval time;
#else
	struct {
		union {
			__u32 tv_sec __attribute__((deprecated));
			__u32 tv_sec_monotonic;
		};
		__s32 tv_usec;
	} time;
#endif
        __u16 type;
        __u16 code;
        __s32 value;
};

Another option is to do a runtime change, and always set the time field
to zero when the kernel is built without support for 32-bit time_t
and user space has not yet called the ioctl to change to monotonic time.

We can also do a combination of the two.

	Arnd
--
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