On Wed, 2011-06-08 at 18:36 +0200, Oleg Nesterov wrote:
> On 06/07, Eric Paris wrote:
> > On Tue, 2011-06-07 at 19:19 +0200, Oleg Nesterov wrote:
> > >
> > > With or without this patch, can't we call audit_syscall_exit() twice
> > > if there is something else in _TIF_WORK_SYSCALL_EXIT mask apart from
> > > SYSCALL_AUDIT ? First time it is called from asm, then from
> > > syscall_trace_leave(), no?
> > >
> > > For example. The task has TIF_SYSCALL_AUDIT and nothing else, it does
> > > system_call->auditsys->system_call_fastpath. What if it gets, say,
> > > TIF_SYSCALL_TRACE before ret_from_sys_call?
> > No harm is done calling twice. The first call will do the real work and
> > cleanup. It will set a flag in the audit data that the work has been
> > done (in_syscall == 0) thus the second call will then not do any real
> > work and won't have anything to clean up.
> Hmm... and I assume context->previous != NULL is not possible on x86_64.
> OK, thanks.
> And I guess, all CONFIG_AUDITSYSCALL code in entry.S is only needed to
> microoptimize the case when TIF_SYSCALL_AUDIT is the only reason for the
> slow path. I wonder if it really makes the measureble difference...
All I know is what Roland put in the changelog:
Avoiding the iret return path when syscall audit is enabled helps
performance a lot.
I believe this was a result of Fedora starting auditd by default and
then Linus bitching about how slow a null syscall in a tight loop was.
It was an optimization for a microbenchmark. How much it affects things
on a real syscall that does real work is probably going to be determined
by how much work is done in the syscall. (or just disable auditd in