Date: Thu, 11 May 2000 09:26:59 -0400 (EDT)
Message-Id: <200005111326.JAA13160@indy.delorie.com>
From: Eli Zaretskii <eliz AT delorie DOT com>
To: Martin Str|mberg <ams AT ludd DOT luth DOT se>
CC: djgpp-workers AT delorie DOT com
In-reply-to: <200005101802.UAA25956@father.ludd.luth.se> (message from Martin
	Str|mberg on Wed, 10 May 2000 20:02:24 +0200 (MET DST))
Subject: Re: Perfomance of gc-simple
References:  <200005101802 DOT UAA25956 AT father DOT ludd DOT luth DOT se>
Reply-To: djgpp-workers AT delorie DOT com
Errors-To: nobody AT delorie DOT com
X-Mailing-List: djgpp-workers AT delorie DOT com
X-Unsubscribes-To: listserv AT delorie DOT com
Precedence: bulk

> From: Martin Str|mberg <ams AT ludd DOT luth DOT se>
> Date: Wed, 10 May 2000 20:02:24 +0200 (MET DST)
> 
> Ok, I tried it on grep. Alas the configure script doesn't find my
> mmap, as the MAP_FIXED option can't be supported, but the test program
> uses it.
> 
> Advice?

It seems that MAP_FIXED is used when the address passed to mmap is
already allocated by some prior call.  (The test program in the
configure script uses malloc to allocate that buffer before it calls
mmap.)  Can someone confirm if this is true in general?

If malloc is always called before mmap(...MAP_FIXED...), then you will
need to scan the data structures maintained by malloc and find the
buffer whose address includes the address passed to mmap.  Then use
that address, after filling it with file's data.

This will obviously lose if the size passed to mmap is larger than the
initial allocation via malloc.  Hmm.  It's possible that MAP_FIXED
cannot be supported in such a case; if many users of mmap need
MAP_FIXED, perhaps this is a show stopper...

Btw, how is this implemented on Unix/Linux?  Here's the relevant
fragment from the test program run by Grep's configure script:

    fd = open("conftestmmap", O_RDWR);
    if (fd < 0)
	exit(1);
    data2 = malloc(2 * pagesize);
    if (!data2)
	exit(1);
    data2 += (pagesize - ((int) data2 & (pagesize - 1))) & (pagesize - 1);
    if (data2 != mmap(data2, pagesize, PROT_READ | PROT_WRITE,
		      MAP_PRIVATE | MAP_FIXED, fd, 0L))
	exit(1);

This expects mmap to return the same address as malloc did (modulo the
alignment).  Since, as DJ explained, mmap uses specially-mapped memory
that can be grown at will without relocating, does this mean that mmap
has to fiddle with the page tables to move the buffer to another
physical address while keeping its logical address unchanged?