Re: Building git-1.5.3.7 on HP-UX 11.00

From: Shawn O. Pearce <spearce@spearce.org>
Date: Dec 11, 2007, 08:54 UTC
Message-ID: <20071211085440.GK14735@spearce.org>
In-Reply-To: <7vfxy9fxh3.fsf@gitster.siamese.dyndns.org>

Junio C Hamano <gitster@pobox.com> wrote:

Show 12 quoted lines

> "H.Merijn Brand" <h.m.brand@xs4all.nl> writes:
> 
> > 3 I am willing to believe that HP_UX' vsnprintf () is broken, or
> >   at least does not conform to the expectations in the GNU world,
> >   but chickening out like the way strbuf_addf () does is maybe a
> >   bit too rude, so I forced a minimum of 64bytes available. That
> >   fixes a lot!
> >
> >   but it still breaks t4013 :(
> 
> I think Shawn had something similar for Solaris, and if it is a small
> cost to help portability, it does not feel so bad.

My Solaris change is already in your tree; on Solaris vnsprintf will return -1 if the input buffer we gave it had a length of 0. This can happen at odd times, depending on what is already in the strbuf and how its buffer is. :)

The fix for Solaris turned out to be just make sure there is at
least 1 byte in the buffer before calling vsnprintf().

> > + git-index-pack -o tmp.idx test-2-7f8ead892057e78576c0329a70cc83afb113f117.pack
> > fatal: serious inflate inconsistency
> 
> That sounds like a broken zlib X-<.

I think this is actually a broken pread(). Its been reported multiple times by different people on HP-UX. Compiling with NO_PREAD=1 resolves the issues for everyone who has tried it.

Jakub asked me for an autoconf test to look for this sort of breakage in pread(), I have yet to come up with a theory for what sort of breakage HP-UX is showing us here, let alone write a test for it.

-- 
Shawn.