The other option is to do some magic to detect the sbrk() calls
which actually try to move the break so that it's aligned,
and ignore them. (And still implicitly page align the initial address).

Jaromir

Emmanuel Dreyfus wrote:
> Hello everybody
> 
> With the help of Kevin B. Hendricks of the Blackdown team, I'm trying to
> fix the way Linux brk() is emulated, so that Linux's JDK-1.3 will work
> in emulation on the PowerPC.
> 
> We use a sample program that reproduces the problem occuring in the JVM.
> The program is basically doing two sbrk(), the second being adjusted so
> that it ends up on a page boundary. [A sbrk() library call result in one
> brk(0) system call to get the brak value, and another brk() system call
> to set the new value with the appropriate offset. Is that right?]
> 
> Here is a Linux emulated by NetBSD trace:
> brk(0)          = 0x10011000
> brk(0x10011021) = 0x10011021
> brk(0)          = 0x10012000
> brk(0x12012fdf) = 0x12012fdf
> 
> and here is the native Linux counterpart:
> brk(0)          = 0x100109d8
> brk(0x100109f9) = 0x100109f9
> brk(0)          = 0x100109f9
> brk(0x20021000) = 0x20021000
> 
> The problem in the emulated trace is that the break value moves between
> the two sbrk(). After setting the break, we end up with 0x10011021 (on
> the second line), and when we read it again, we get 0x1001200 (on the
> third line). And the adjustment fails. Here is the explanation with more
> details
> 
> The program first calls sbrk(), it wants to allocate some memory.
> 
> brk(0)          = 0x10011000
> brk(0x10011021) = 0x10011021
> 
> The program want to add a 0xfdf so that we reach a page alligned
> address: 0x10011021 + 0xfdf = 0x10012000. It calls sbrk() again.
> 
> sbrk() calls brk(0) to get the break address
> 
> brk(0)          = 0x10012000
> 
> But it's not at 0x10011021 anymore, anyway, the library call sbrk() does
> not care, it has just been asked to add 0xfdf.
> 
> brk(0x12012fdf) = 0x12012fdf
> 
> And the result is not page aligned, which causes the JDK to fail.
> 
> The problem is caused by the linux_sys_brk(), in
> /sys/compat/linux/common/linux_misc.c. And it is annotated as being a
> problem:
> 
>    oldbrk = vm->vm_daddr + ctob(vm->vm_dsize);
>    /*
>     * XXX inconsistent.. Linux always returns at least the old
>     * brk value, but it will be page-aligned if this fails, 
>     * and possibly not page aligned if it succeeds (the user 
>     * supplied pointer is returned).
>     */
>    SCARG(&oba, nsize) = nbrk;
>  
>    if ((caddr_t) nbrk > vm->vm_daddr && sys_obreak(p, &oba, retval) ==
> 0)
>       retval[0] = (register_t)nbrk;
>    else
>       retval[0] = (register_t)oldbrk;
>  
>    return 0;
> 
> I need some information about this ctob(vm->vm_dsize). What is it for?
> As I understood vm->vm_daddr is the current break value, why do we add
> something to it?
> 
> -- 
> Emmanuel Dreyfus.  
> Avec Windows 3.1 ils etaient au bord du gouffre...
> Avec Windows 95 ils ont fait un grand bon en avant.
> p99dreyf@criens.u-psud.fr    
> 


-- 
Jaromir Dolecek <jdolecek@NetBSD.org>      http://www.ics.muni.cz/~dolecek/
 ***  Wanna a real operating system ? Go and get NetBSD, dammit!  ***