Re: 2.6.12-rc6-mm1

[Wolfgang Wander]

Wolfgang Wander wrote:
> Andrew Morton wrote:
>
> > +avoiding-mmap-fragmentation-revert-unneeded-64-bit-changes-vs-x86_64-task_size-fixes-for-compatibility-mode-processes.patch 
>
>
> As a heads-up.
>
> This one breaks the fragmentation reduction patch in 32 bit emulation mode.
> Our test case shows the standard 17 fragmented regions in 
> /proc/self/maps (as in
> the 2.6 standard kernel) vs the 2 regions in 2.6.12-rc5-mm2 (and before).
>
> Somehow the new way of detecting 32 bit remulation mode seems to fail here.
>
> I'll try to figure out a fix.

Here is one possibility:

Since rc6 the difference between TASK_UNMAPPED_64 and TASK_UNMAPPED_32 is gone
and both are now merged into TASK_UNMAPPED_BASE.  Therefore we can no longer
check our local base against TASK_UNMAPPED_BASE to see if we are running in 32bit
emulation mode.  The appended patch uses other (hopefully the right) means.

Tested on x86_64 in 32 and 64 mode (64 bit fragments as desired, 32 bit
collapses as desired).

Signed-off-by: Wolfgang Wander <wwc@xxxxxxxxxx>

--- arch/x86_64/kernel/sys_x86_64.c~	2005-06-07 09:12:31.000000000 -0400
+++ arch/x86_64/kernel/sys_x86_64.c	2005-06-07 10:32:07.000000000 -0400
@@ -105,7 +105,8 @@ arch_get_unmapped_area(struct file *filp
 		    (!vma || addr + len <= vma->vm_start))
 			return addr;
 	}
-	if (begin != TASK_UNMAPPED_BASE && len <= mm->cached_hole_size) {
+	if (((flags & MAP_32BIT) || test_thread_flag(TIF_IA32))
+	    && len <= mm->cached_hole_size) {
 	        mm->cached_hole_size = 0;
 		mm->free_area_cache = begin;
 	}