Re: [PATCH v2 1/2] Y2038: make __mktime_internal compatible with __time64_t

[Lukasz Majewski <lukma at denx dot de>]

Hi Paul,

> On 3/28/19 1:59 AM, Lukasz Majewski wrote:
> > In other words - I shall not introduce new "installed" type for
> > struct timeval and just in posix/bits/types.h define:
> >
> > #ifndef __USE_TIME_BITS64
> >   __STD_TYPE __TIME_T_TYPE __time_t; /* Seconds since the Epoch. */
> > #else
> >   __STD_TYPE·__TIME64_T_TYPE __time_t;
> > #endif
> >
> > In that way all structures which use __time_t are Y2038 safe.
> >
> > ONE NOTE:
> > I also guess that the above change keeps those structs posix
> > compliant for 32 bit machines ?  
> 
> Yes, that's the idea.
> 
> 
> >  
> >> However, we should
> >> not support a complicated API like that, as it's typically not
> >> useful in practice and its mere availability causes more confusion
> >> than it's worth - as I've discovered with _FILE_OFFSET_BITS and
> >> __off64_t.  
> >
> > If I may ask - what were the biggest problems?  
> 
> The biggest problem is confusion and complexity. For example, see
> /usr/include/zlib.h (assuming you have zlib installed). zlib is one of
> the few libraries that tries to support both 32- and 64-bit file
> offsets in user code. Almost nobody understands how it is supposed to
> work, and I'm not sure that it even does work. And even in code that
> doesn't attempt to support this sort of thing, there are still
> problems. See, for example, the comedy of errors in
> <https://stackoverflow.com/questions/22663897/unknown-type-name-off64-t>
> where people suggest monstrosities like -Doff64_t=_off64_t to work
> around compilation issues with the Apache Portable Runtime.
> 
> Rather than go down those rabbit holes, it's better to say that the
> entire program must be compiled consistently, i.e., one must compile
> all libraries with -D_TIME_BITS=64 if one plans to use
> -D_TIME_BITS=64 in user code that passes time_t to these libraries.
> This includes OpenSSL, GnuTLS, Kerberos, Glib/Gtk, libpng,
> ImageMagick, etc., etc., as all these libraries expose time_t in
> their APIs. And it's not just libraries: for example, one should
> compile Python with -D_TIME_BITS=64 and all Python modules requiring
> native code and using time_t in their API should also be built with
> -D_TIME_BITS=64.
> 
> Although it'll be a real hassle to insist on -D_TIME_BITS=64
> everywhere, it's the only approach that will really work in practice.

The above approach is doable if building the embedded system rootfs in
OE/Yocto.

I do have a setup where only for glibc compilation I do disable
-D_TIME_BITS64, but for _all_ other source code it is enabled.

> That's life. Frankly if it were up to me, I'd make 64-bit time_t the
> default and ask people to compile with -D_TIME_BITS=32 everywhere if
> they have backward-compability concerns, as this will be much better
> in the long run, and would help us avoid many of the issues we ran
> into during the off_t debacle.
> 

I see your point.

> 
> >> Instead, glibc should have a simple user API where _TIME_BITS=64
> >> merely says "I want 64-bit time_t everywhere" and a module either
> >> uses this option or it doesn't.
> >>  
> > So according to above I shall only introduce glibc _internal_ struct
> > __timespec64/__timeval64 in include/bits/types/ :
> >
> > #if __WORDSIZE > 32 || ! defined(__USE_TIME_BITS64)
> > # define __timespec64 timespec;
> > #else
> > struct __timespec64 {
> > 	__time64_t tv_sec;
> > 	__int64_t tv_nsec;
> > }
> > #endif  
> 
> No, these internal interfaces should not be in any file installed
> under /usr/include where users can find them and get confused. 

Yes, correct.

I've proposed to install struct __timespec64/__timeval64
in ./include/bits/types directory of glibc (and as fair as I understand
those would be private).

But also, those can be placed in ./include/time.h

> They
> should be only in .h files that are private to glibc and are not
> installed. A logical place for them would be include/time.h.
> 

Ok.

> >
> > and rewrite the relevant functions/syscalls (like clock_settime() in
> > this particular case) to use it in glibc ?  
> 
> Yes, implementations of syscalls can use these internal interfaces.
> 

Ok. Good.

> 
> >
> > PROBLEM(S) with internal struct __timespec64:
> >
> > - Would be misleading for 32 bit architectures (minor issue)  
> 
> I'm not sure I understand this point. How would we be misleading users
> by keeping the __time64_t interfaces private?

With __timespec64 exported (installed in usr/include/*) - there would
be an explicit type to show that we are Y2038 safe (when e.g.
debugging).

> 
> 
> >
> > - Needs to met specific Linux kernel's ABI as it is passed as an
> >   argument to Linux syscalls (like clock_settime64).  
> 
> Yes, internally glibc will need to know what the kernel ABI is, and do
> the right thing. This should be reasonably easy to do if the internal
> glibc interfaces are 64-bit, which they should be.

As presented above - kernel expects in e.g. timespec 64 bit values for
both tv_sec and tv_nsec. Also internal types doesn't need to comply with
POSIX (tv_nsec shall be single long as seen from user space).

> 
> 




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: pgpjH12hGLPGO.pgp
Description: OpenPGP digital signature