> The future proof (and POSIX) way of doing ANY restriction of ram allocation
> is the posix typed memory interface.  The old MAP_* flags are kinda hokey,
> and had lots of 'hacks' to make things like this work.
> 
> You are supposed to setup the relevant regions in startup.  Now I've mentioned
> 
> to people before we need to have these defined BEFORE the 6.4 release...
> 

Are ya saying that if I add memory to a machine i need to twiddle with the startup stuff?

About 6.4 don't know if it's close to release or not but it's been a while since there hasn't been a Milestone release (
 last december )

> Time for a PR, methinks...
> 
> Mario Charest wrote:
> >  > Hmmm, that would be POSIX typed memory, I think.
> > 
> > Are you talking about posix_typed_mem_open(). 
> > 
> > So in a nut shell if I want the memory bloc to be below 4G I use mmap, 
> > if 'd want it to be above 4G i would use posix_....  If I don't care, 
> > I'm force to make a choice?
> > 
> > That doesn't sound right.
> > 
> >  >
> >  > Mario Charest wrote:
> >  > > That brings the question how can I force contiguous memory to be
> >  > > allocated above 4G, below 4G or anywhere?
> >  > >
> >  > > - Mario
> >  > >
> >  > > _______________________________________________
> >  > > OSTech
> >  > > http://community.qnx.com/sf/go/post5541
> >  > >
> >  >
> >  > --
> >  > cburgess@qnx.com
> > 
> > 
> > 
> > 
> > _______________________________________________
> > OSTech
> > http://community.qnx.com/sf/go/post5560
> > 
> 
> -- 
> cburgess@qnx.com

On Thu, Mar 06, 2008 at 09:42:37AM -0500, Mario Charest wrote:
> So what you are saying is that 6.4 will break old driver if use with more
then 4G.  

No, it _may_ break. If the driver is requesting contiguous memory it will
likely 
work - 'normal' user allocations are preferentially done from above 4G to
leave 
the memory below for allocations that need it. That means that there will 
normally be memory free below the line.

There's another out - implicit in the memmgr code is ability to have
an individual restriction list for each memory object (a restriction list is

what a typed memory name turns into - a linked list of paddr start/end
ranges 
that the physical allocator is allowed to use for a particular allocation).
It's always been in our plan (but not yet implemented) to have an API that
let you start a process and specify a typed memory name to use for its
non-typed memory mmap()'s (what's happening right now is that all objects
are using the typed memory name "sysram" for their allocations). If the API
was implmented, you could theoretically do something along the line of:

	on -M"sysram_below_4G" old_driver

and all the mmap()'s that the process does would come from below the 4G
line.
A little bit of overkill since not all the memory the driver uses is
required
to be below the line, but better than nothing.  Unfortunately, we're not 
likely to have the time to do that before the 6.4 release.	

> I though the OS could make the assumption that if mmap is used with
MAP_PHYS 
> and MAP_ANON that it could only then allocate it below 4G.

As I said previously, we can't make that assumption. The memory manager
code has no knowledge of if the memory is even going to be handed off to
a device or what the restrictions on the device memory usage are.

-- 
Brian Stecher (bstecher@qnx.com)        QNX Software Systems
phone: +1 (613) 591-0931 (voice)        175 Terence Matthews Cr.
       +1 (613) 591-3579 (fax)          Kanata, Ontario, Canada K2M 1W8