nss and pthreads and wine, oh my!

[Daniel Eischen]

On Tue, 9 Dec 2003, Craig Boston wrote:

> I've come across a problem that I believe had been plaguing me on -CURRENT for 
> a while but I just now was able to track down a possible cause.  The symptoms 
> are that wine is unable to run any Win32 program that uses gethostbyname or 
> getaddrinfo.  I'm seeing this on a recent 5.2-BETA and also an older 
> 5.1-RELEASE, so I can rule out the statfs ABI changes.  Wine is built from 
> source on both and without CPUTYPE on the 5.2-BETA machine.
> 
> After building wine (and libc) with debug symbols, I can reproduce the 
> following backtrace of where the crash occurs:
> 
> #0  0x5ce20000 in ?? ()
> #1  0x5c0f3c98 in _nsdispatch (retval=0x5c71f054, disp_tab=0x5c131400,
>     database=0x5c12c920 "hosts", method_name=0x5c12c912 "gethostbyname",
>     defaults=0x5c1313d0) at /usr/src/lib/libc/net/nsdispatch.c:574
> #2  0x5c0f3004 in gethostbyname2 (name=0x5d031c40 "test",
>     type=2) at /usr/src/lib/libc/net/gethostnamadr.c:88
> #3  0x5c0f2f9e in gethostbyname (name=0x5d031c40 "test")
>     at /usr/src/lib/libc/net/gethostnamadr.c:72
> 
> line 574 of nsdispatch.c looks like this:
> 
> 574             result = _pthread_rwlock_rdlock(&nss_lock);
> 
> Now, I've never really delved into libc before, so everything from this point 
> is pure guesswork.  However, it strikes me as odd that nsdispatch() is 
> calling pthread locking primitives in the non-reentrant C library (wine is 
> linked against libc.so.5).
> 
> A quick grep of the source code shows that the only other thing similar in 
> libc is in the RPC code, but in that case it uses a macro that checks 
> __isthreaded before calling the pthread function.  According to the comments, 
> __isthreaded is only set if a process has more than 1 thread.  So, I tried 
> wrapping all of the pthreads stuff in nsdispatch.c with if(__isthreaded) 
> blocks and presto, wine is working again.
> 
> I have no idea why wine seems to be the only program that behaves this way...  
> At this point I'm not sure if it's a bug in libc or a bug in wine.  Has 
> anybody else seen similar behavior out of wine or anything else?
> 
> Are there any threads gurus out there who know what the 'right thing' for 
> nsdispatch to do is?

Yes, I think libc should only call locking primitives when
__isthread != 0.

I am a bit confused though, because I would think nsdispatch works
for other cases and don't know why it isn't for wine.  What thread
library are you using?

-- 
Dan Eischen