Re: [RFC] de-asmify the x86-64 system call slowpath

[Oleg Nesterov <oleg@redhat.com>]

On 01/28, Al Viro wrote:
>
> On Tue, Jan 28, 2014 at 05:38:08PM +0100, Oleg Nesterov wrote:
> > On 01/28, Al Viro wrote:
> > >
> > > On Mon, Jan 27, 2014 at 06:39:31PM +0100, Oleg Nesterov wrote:
> > > > On 01/27, Al Viro wrote:
> > > > >
> > > > > Why is _TIF_UPROBE *not* a part
> > > > > of _TIF_DO_NOTIFY_MASK, for example?
> > > >
> > > > Yes, please see another email. That is why uprobe_deny_signal()
> > > > sets TIF_NOTIFY_RESUME along with TIF_UPROBE.
> > >
> > > *grumble* Can it end up modifying *regs?  From very cursory reading of
> > > kernel/events/uprobe.c it seems to do so, so we probably want to leave
> > > via iretq if that has hit, right?
> >
> > But we do this anyway, restore_args path does iretq?
> >
> > I mean, uprobe_notify_resume() is called from do_notify_resume(), it
> > should be fine to modify*regs there?
>
> See Linus' patch trying to avoid iretq path; it's really costly.  Looks
> like that patch will have to treat _TIF_UPROBE the same way it treats
> _TIF_SIGPENDING...

Ah, this one I guess: http://marc.info/?l=linux-kernel&m=139077532507926

I think this should be fine wrt uprobes, unless I misread this patch
syscall_exit_slowpath() is actually only called by ret_from_sys_call
path, it this case TIF_UPROBE should not be set. But perhaps
"retval = 1" after uprobe_notify_resume() makes sense anyway.

And while I am almost sure I missed something, can't we (with or without
that patch) simply add TIF_UPROBE into _TIF_DO_NOTIFY_MASK and remove
set_tsk_thread_flag(t, TIF_NOTIFY_RESUME) from uprobe_deny_signal ?

Oleg.