Re: Threading behavior in 2.6.5 may be broken ?

[Joe Korty <joe.korty@ccur.com>]

On Wed, Jun 09, 2004 at 09:16:41PM -0700, j.random.programmer wrote:
> Hi all:
> 
> I just installed Fedora Core 2 (2.6.5.x smp 
> kernel) on a Dual 1 Ghz P4 server with about 
> 1.5 GB of RAM and about 1.4 GB of swap. 
> 
> I am primarily a web/database developer, not 
> a C programmer so I am writing this email from
> an end-user's perspective.
> 
> I have a program that tries to create as many
> threads as possible. This program was written by me 
> for kicks/testing -- just to see what would happen.
> I ssh into the server and run this program as root
> under the sun 1.4.2 JDK.
> 
> On a 2.4.x kernel, from a Java JVM I could create
> about 900 threads before the JVM crapped out with
> a "cannot create more threads" type of error. Before
> this point, I can create/run - say 700 threads - just
> fine. This is good -- a clean failure at some point
> and good behavior before then.
> 
> On this new kernel, the system gets totally wedged
> when I run the same program and try to create
> 10,000 threads. Instead of getting a "cannot
> create more threads" error, I now get an "out of
> memory" error. Then the command line freezes in
> the existing terminal window, ctrl+c does not work
> (no matter how many times it's pressed), I cannot
> launch another ssh session and cannot ssh into the
> server again (although ping still works).
> 
> To recap:
> 
> [2.4.x]
> 700 threads --> fine
> 10,000 threads --> crap out at 900 something. 
> 
> [2.6.5]
> 700 threads --> fine
> 10,000 threads --> system wedged totally. 

It's funny you should post this, I just encountered the
same thing a few days ago, and found the cause a few
minutes ago.

The problem is that 2.6.x restricts unpinned mmaps (mmaps
where the app does not supply the virtual address) to the
400000000-c000000000 range, while redhat allows all holes
in the user address space to satisfy such requests.

RedHat implements this by introducing an i386-specific
version of arch_get_unmapped_area().  This has not
yet made it to the official kernels.  I have no idea
if it was ever submitted for consideration, or was
submitted and rejected.

Regards,
Joe