mmaping a fixed address fails on ARM

[Mikael Pettersson]

Bryan Wu writes:
 > Let me add Andre here
 > 
 > Thanks,
 > -Bryan
 > 
 > On Thu, Jan 13, 2011 at 5:58 AM, Bryan Wu <bryan.wu at canonical.com> wrote:
 > > Hi Russell,
 > >
 > > Andre posted a mmap testcase [1] and a bug report [2] for Ubuntu
 > > kernel on OMAP4 system, since he is porting some applications from x86
 > > to ARM. He testcase works fine on x86, but always fail on ARM. I've
 > > tested it on OMAP3/OMAP4/i.MX51.
 > >
 > > Basically what he want to do is get a mapping on a specific virtual
 > > address. But AFAIK, mmap doesn't make sure we can get the mapping
 > > address as we want. On x86, it works fine. So if this fails, the
 > > application can't run on ARM.
 > >
 > > The implementation of arch_get_unmapped_area() is different from ARM
 > > and x86, that might makes mmap behavior different between ARM and x86.
 > > Andre said 2.6.28-versatile. I'm not sure whether it is a regression.
 > >
 > > [1]: http://launchpadlibrarian.net/61588086/main.c
 > > [2]: https://bugs.launchpad.net/ubuntu/+source/linux-ti-omap4/+bug/697004

mmap() with a non-NULL hint referring to a currently unmapped area,
but without MAP_FIXED, isn't even remotely portable or reliable
(I've tried it myself in the context of JITs).  It may work on
Linux/x86, but that doesn't allow you to assume it should work
on other OS/CPU combinations (and I know of several where it
emphatically doesn't work).

If you want a fixed address then use MAP_FIXED.  If you don't want
MAP_FIXED to clobber an existing mapping then you have to keep track
of existing mappings yourself and avoid them.

What is that "some application"?

/Mikael