Re: Accelerating Y2038 glibc fixes

[Lukasz Majewski <lukma at denx dot de>]

Hi Zack,

> On Fri, Jul 26, 2019 at 6:39 AM Lukasz Majewski <lukma@denx.de> wrote:
> ...
> > > > See for example [1] - there are just 7 lines of "code".  But
> > > > Joseph does not accept our patches.  The arguments he gives are
> > > > not on a technical level;  
> ...
> > Our goal is to add a solid foundation for the Y2038 work, so we
> > would know the direction where we are heading.  
> ...
> > If you think that it would be better and most of all faster if you
> > rewrite the description, then I don't mind.
> >
> > It would be great if you could do it sooner than latter as this
> > slows down our development.  
> ...
> > The most recent version of this patch set (v8):
> > https://patchwork.ozlabs.org/patch/1117100/  
> 
> I haven’t been following the details of these patches super carefully,
> and I’m not sure I understand what _Joseph’s_ concerns with your
> writing is.  However, I’m a native English speaker, I’ve read over the
> text in the patch at <https://patchwork.ozlabs.org/patch/1117100/>, I
> do think I understand the issues at a high level, and I do think the
> meaning of __ASSUME_TIME64_SYSCALLS could be explained more clearly.
> I’m prepared to work with you to come up with better wording 

Thanks for offering your help.

> but I
> need to ask you a bunch of questions.  Could you please reply to each
> of the queries marked Qn below?
> 
> As I understand it, we have five distinct cases to consider:
> 
> 1. All future LP64 ABIs and all but one existing LP64 ABI, identified
>    by __WORDSIZE == 64: time_t is already a 64-bit integer type and
>    all of the relevant system calls already accept it in that form.
>    glibc’s implementation of, for instance, clock_gettime may continue
>    to invoke the system call numbered __NR_clock_gettime.

This is exactly how we shall proceed with machines having
__WORDSIZE==64 (e.g. x86_64, armv8, etc).

They now support 64 bit time with non suffixed syscalls.

In the patch [1] the __WORDSIZE == 64 check covers this.

> 
> 2. The exception to (1) is Alpha.  That is an exclusively LP64
>    architecture, but in glibc 2.0 it used 32-bit time_t, and we still
>    have compatibility code for that case.  The compatibility symbols
>    invoke a set of compatibility syscalls with ‘osf’ in their names:
>    for instance, gettimeofday@GLIBC_2.0 invokes __NR_osf_gettimeofday.
>    Not all of the time-related functions in glibc have compatibility
>    symbols, only those that existed in version 2.0.
> 
>    Your patches do not touch this compatibility code at all, as far as
>    I can see. 

Yes, you are correct. I was not even aware of such a case (as I found
Alpha as 64 bit arch when I did my checking).

> Alpha being out of production, and binaries compiled
>    against glibc 2.0 being rare anyway, it would only make sense to
>    involve this code in your patches if it reduced the overall volume
>    of compatibility code somehow, but regardless we need to make sure
>    it doesn't break.

As you mentioned - we shall not break existing binaries. However, I'm
not sure if we shall spent more time/effort on the arch being near EOL
(or at least being out of production now).

> 
> 3. x32 (recently new 32-bit ABI for x86): like case 1, time_t is
>    already 64-bit and we use unsuffixed system calls.  The text says
>    this case is identified by __WORDSIZE == 32 && __TIMESIZE == 64,
>    but the code actually checks __SYSCALL_WORDSIZE.
> 
>    Q1: Which condition correctly identifies this case, __TIMESIZE ==
> 64 or __SYSCALL_WORDSIZE == 64?

It is:

(__WORDSIZE == 32) && ((defined __SYSCALL_WORDSIZE &&__SYSCALL_WORDSIZE
== 64))

Only x32 defines the __SYSCALL_WORDSIZE == 64 (as it has __WORDSIZE ==
32, but supports 64 bit syscalls ABI).

> 
>    Q2: Could we perhaps ensure that __TIMESIZE and/or
> __SYSCALL_WORDSIZE is defined to 64 whenever __WORDSIZE == 64? Then
> we could collapse this into case 1.

__TIMESIZE == 64 for x32. 

The x32 uses the same set of syscalls (e.g. clock_gettime) as in point 1
(as for example x86_64).

> 
> 4. Brand-new (added in kernel 5.1 or later) 32-bit ABIs: time_t will
>    always be 64-bit,

This would be true after we make the "switch" to support Y2038 aware
systems. Please find example implementation [2] from this patch series
[3] (it adds example code for converting __clock_settime to support 64
bit time when __ASSUME_TIME64_SYSCALLS is defined).

> _but_ glibc’s implementation of time-related APIs
>    may need to invoke system calls with suffixed names: clock_gettime
>    invoking __NR_clock_gettime64, for instance.  Also, the kernel will
>    not provide all of the time-related system calls that have
>    historically existed; glibc must, for instance, implement
>    gettimeofday in terms of clock_gettime.

Yes, correct. Some syscalls would be emulated (as they are not or will
not be converted to 64 bit version).

> 
>    Q3: What macros are defined for this case?

There are no macros yet available.

> 
>    Q4: Does glibc need to call system calls with suffixed names in
>    this case?

I think yes - for example the gettimeofday would internally use
clock_gettime64 (vDSO if available).

> 
>    Q4a: If the answer to Q4 is ‘yes’, why is that, and can we change
>    the kernel so that it’s the same as x32 and the LP64 architectures?

We need new set of syscalls for 64 bit time support on 32 bit archs
(WORDSIZE==32); for example x32/LP64 would still use clock_settime
syscall (number X). To have the same functionality (64 bit time
support) 32 bit archs would need to use clock_settime64 (number 404 on
armv7)

And here the __ASSUME_TIME64_SYSCALLS comes into play. If the arch is
capable of providing 64 bit time, (no matter if it uses clock_settime
or clock_settime64), then __ASSUME_TIME64_SYSCALLS is defined.

It is also assumed that both clock_settime64 and clock_settime provide
the same ABI, so no code needs to be adjusted in glibc.

If code needs to be adjusted (as the calls are not compatible) - a new
flag will be introduced (like with semtimedop)

>    (Either by removing the suffixes, or by _adding_ suffixed aliases
>    to asm/unistd.h for x32 and LP64 architectures.)

Wouldn't this caused the ABI break?

> 
> 5. Historical 32-bit ABIs, where the existing set of system calls
>    takes 32-bit time_t, and Linux 5.1 added a matching set that takes
>    64-bit time_t.  For compatibility with old programs that make
>    direct system calls, the kernel will not rename the __NR_ constants
>    for the old (32-bit) system calls; instead new constants with ‘64’
>    or ‘time64’ suffixes will be added.  As in case 4, the new set of
>    system calls does not cover all of the historic time-related system
>    calls.
> 
>    In this case, and only this case, glibc’s code needs to account for
>    the possibility that the new __NR_ constants are not defined
>    (because glibc is being compiled against kernel headers from a
>    version older than 5.1), or that the new system calls are not
>    available at runtime (glibc was compiled against new kernel headers
>    but is running with an old kernel).
> 
>    The #if is complicated enough that I’m not sure, but I _think_ your
>    code only defines __ASSUME_TIME64_SYSCALLS when the new constants
>    are _guaranteed_ to be defined.
> 
>    Q5: Is it correct that __ASSUME_TIME64_SYSCALLS is only defined
>    when the new constants are guaranteed to be defined?

No.

The __ASSUME_TIME64_SYSCALLS is defined only when the architecture
supports 64 bit time related ABI.

(either via clock_settime on e.g. x86_64/arm64 or clock_settime64 on
arm).

Please consult the code for clock_settime [4]. It shows how the
__ASSUME_TIME64_SYSCALLS flag is used in practice.

> 
>    Q6: All of the other __ASSUME_ constants mean both that we assume
>    the kernel headers are new enough to provide all the necessary
>    declarations, and that we assume the feature is available at
>    runtime: no fallback code will be included in the library.  Is this
>    also the intended meaning of __ASSUME_TIME64_SYSCALLS?

The patch [1] defines the __ASSUME_TIME64_SYSCALLS as the ability of
the architecture (via kernel syscalls) to provide 64 bit time support.

As shown in [4] - the fallback code is called when
__ASSUME_TIME64_SYSCALLS is NOT defined and if architecture doesn't
support clock_settime64.

> 
> zw


Note:

[1] - https://patchwork.ozlabs.org/patch/1117100/ 

[2] -
https://github.com/lmajewski/y2038_glibc/commit/3d5f3512438de7426acba58c1edf53f756f8570b#diff-c051022b496f12bd9028edb46b8ec04d

[3] -
https://github.com/lmajewski/y2038_glibc/commits/Y2038-2.29-glibc-__clock-internal-struct-timespec-v6

[4] -
https://github.com/lmajewski/y2038_glibc/commit/69f842a8519ca13ed11fab0ff1bcc6fa1a524192



Best regards,

Lukasz Majewski

--

DENX Software Engineering GmbH,      Managing Director: Wolfgang Denk
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-59 Fax: (+49)-8142-66989-80 Email: lukma@denx.de

Attachment: pgpJCjyujenXT.pgp
Description: OpenPGP digital signature