Re: why mprotect(PROT_SEM) ?

"Dr. Uwe Girlich" schrieb:
> 
> Hello list!
> 
> In my quest to remove (or at least handle the size dynamically) of
> /dev/shm/ngpt I just read about sys_futex.c/sys_futex_region() and found, it
> calls in the end only
>         mprotect(area, size, PROT_READ|PROT_WRITE|PROT_SEM);
> The read and write protection should be OK anyway (the file /dev/shm/ngpt
> itself has access rights 666) so the only interesting point here is PROT_SEM.
> This macro came new into the kernel with the futex patch. It is defined in
> asm-(IBM-architectures)/mman.h as
> 
> #define PROT_SEM 0x8 /* page may be used for atomic ops */
> 
> and used only once: in mm/mprotect.c/sys_mprotect() this new flag is also
> allowed but nobody does anything with it.
> 
> So what can we learn from this? Is this mprotect() call necessary at all? May
> it be, that other architectures may need to prepare something more to allow
> atomic operations in user pages?
> 

In 2.4 the mprotect is necessary to prevent that the page gets evicted
and paged in at different physical address. The physical address and
offset is used as a hash for the futex.

In 2.5 we have now a cache for mappings virtual -> physical with a callback
notification.

I hope that makes sense to you.