Re: PCH support on Solaris

["Bonzini" <bonzini at gnu dot org>]

> > gives the same address for both guess and size?  That is, that Solaris
> > completely ignores the first argument unless MAP_FIXED is given?
>
> It appears so, indeed.
>
> > If so, my idea is useless at least under Solaris.
>
> No, it is not useless under Solaris

It's me who's lost now. :-)  If I understand it correctly, you are saying that on one
hand giving mmi.preferred_base to the /dev/zero memory mapping is useless, on the other
hand it "just works".  This is playing with fire about as much as the current, not
working, code; you should really use Ulrich's idea.  You choose whether to embed my
heuristic in the following snippet.

#ifdef HAVE_MINCORE
    /* We can make sure that MAP_FIXED is safe */
    int flags = MAP_FIXED|MAP_SHARED;
    char *vec = alloca ((mmi.size + pagesize - 1) / pagesize);
    int result = mincore (mmi.preferred_base, mmi.size, vec);
    if (result == -1)
       {
            warning ("PCH not usable");
            return;
        }
#else
    /* We cannot make sure that MAP_FIXED is safe, just
        use some common sense.  It might be possible to try
        to test the availability of a region at mmi.preferred_base
        by using a non-fixed anonymous mmap, but this is not
        bulletproof because the OS might just choose to ignore
        the first argument when MAP_FIXED is not used. */
    int flags = MAP_SHARED;
#endif

addr = mmap (mmi.preferred_base, mmi.size,
    PROT_READ, flags, fd, mmi.offset);

if (addr != mmi.preferred_base)
    {
        assert (addr == -1);
        if (addr == (void *) -1)
            perror ("mmaping precompiled header");
        warning ("PCH not usable");
        return;
    }

/* ok, mmap done */

|_  _  _ __
|_)(_)| ),'
------- '---