[lkml]   [2001]   [Jul]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: increasing the TASK_SIZE
On Thu, Jul 05, 2001 at 05:21:21PM -0400, Ernest N. Mamikonyan wrote:
> I was wondering how I can increase the process address space, TASK_SIZE
> (PAGE_OFFSET), in the current kernel. It looks like the 3 GB value is
> hardcoded in a couple of places and is thus not trivial to alter. Is
> there any good reason to limit this value at all, why not just have it
> be the same as the max addressable space (64 GB)? We have an ix86 SMP
> box with 4 GB of RAM and want to be able to allocate all of it to a
> single program (physics simulation). I would greatly appreciate any help
> on this.

It is marginally possible to increase that up so much
that you get about 3.8-3.9 GB for usermode process.
(I use k=1024, M=k*k, G=k*k*k)

It is absolutely impossible to get it into anything above
the 4.0 GB limit. This hard limit is buried inside the i386
(and all of its successors) memory addressing, and mapping
hardware. There is a choke-point of 32 address bits along
the way, which prevents going above 4.0 GB most effectively.

With considerable infrastructural work(*) it MIGHT be possible
to go very near the 4.0 GB limit for userspace, but I am not
an expert here. The crux is at the supervisor/interrupt mode
stack memory mapping. As far as I understand, in i386 we
must have the supervisor stack (and 'struct task') mapped
into the same address space as the usermode. Only the memory
protection prevents the usermode to access that data.
Also parts of kernel code must be in that address space + parts
of kernel data related into MMU control.

(*) Supervisor (kernel) mode must have the stack, and switch-
around code + some datasets in its access space when transition
into the kernel space is done (and reversed). Accessing user-
space from kernel can then be done via kmap() (-like) windows.
Of course this is considerably much slower than the current method
where each user-space has 1/4 of its total address space allocated
for kernel internal use.

To get most out of your box, you need to run your problem as much
as possible at separate processors and with separate contexts.
That way you will get most out of your SMP setup.
(Consider your box as a small Beowulf-cluster.)

Of course problems where you run e.g. PVM, you will need fast
communication in between processes, and nothing would beat single
shared memory space. You might be able to get that by having
e.g. SHM segments used for PVM's IPC task.
Linux doesn't support user semaphores in SHM in scheduling sense,
though. You can, of course, do CPU burning spin-locks for shared
memory area access. The best would, IMO, be a hybride of using
SHM for transfering large amounts of data in between processes,
and something alike PF_UNIX sockets for signaling that there is
some new data available.

In _usual_ case you can ignore such details, and use your favourite
clustering library, like PVM.

> Thanks a great deal,
> Ernie
> PS. Please `CC' me the answer!
> Ernest N. Mamikonyan E-Mail:
> Philadelphia, PA 19104 Web:

/Matti Aarnio
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2005-03-22 12:56    [W:2.205 / U:0.116 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site