On Sat, Mar 10, 2012 at 3:25 PM, Djalal Harouni <tixxdz@xxxxxxxxxx> wrote:
>
> Given that consideration this patch introduces two counters:
> A global atomic execve counter that will be incremented on every
> do_execve_common() call, and an atomic exec_id member for the task_struct.
This seems horribly expensive on most 32-bit architectures, including
very much x86-32. That atomic64_inc_return() is not cheap. It's
possible that it's basically an impossible operation to do atomically
on certain platforms, causing it to use some random spinlock instead.
I wonder if we couldn't make something much cheaper, since we don't
actually care about it being globally incrementing, we just care about
it being globally unique. IOW, it could easily be a 56-bit per-cpu
counter along with the CPU number in the high bits or something like
that. Avoiding the whole atomicity issue, and thus avoiding the main
reason those things are really expensive.
IOW, something like
cpu = get_cpu();
.. increment percpu 64-bit counter ..
id = counter * MAX_CPUS + cpu;
put_cpu(cpu);
or equivalent would seem to be a potentially much cheaper approach.
Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
