libpthread vs libthr.

[Daniel Eischen]

On Sat, 11 Nov 2006, Dag-Erling Smørgrav wrote:

>> 	resolve a symbol, sched_yield(2).  So X server didn't run.
>>
>> 	In this case, gdm tried to resolve a symbol,
>> 	sched_yield at LIBTHREAD_1_0 instead of sched_yield.  So by changing
>> 	libthr by libmap, gdm couldn't resolve a symbol, sched_yield(2).
>> 	libthr doesn't have a sched_yield at LIBTHREAD_1_0, Yes, libc have
>> 	a sched_yield, but sched_yield at LIBTHREAD_1_0.
>
> This is not a libthr issue, it's a symbol versioning issue.  Arguably,
> sched_yield should be removed from libpthread's symbol map, because
> it's a wrapper for a syscall and we don't version syscalls.

Yah we do, see src/lib/libc/sys/Symbol.map.

In the above case(s), libpthread and libthr use LIBTHREAD_1_0 as the
version namespace.  This was done back in 5.x.  But now that libc
has symbol versioning and the thread libraries want to override
some symbols in libc, they (libc and the thread libraries) need
to use the same version namespace.  I added LIBTHREAD_1_0 as a
compatibility hack to libpthread in -current, but it also has
and defaults to FBSD_1_0 which is also used by libc.

The way to fix this for the above problem is to enable symbol
versioning (set SYMVER_ENABLED=yes in /etc/make.conf) and rebuild
world.  Then rebuild your ports.  After that, all your ports will
reference symbols in FBSD_1_0, and if libthr doesn't provide
sched_yield at FBSD_1_0, it'll find it in libc.

Before 7.0 goes out the door, symbol versioning will be enabled
by default so you won't have this problem.

-- 
DE