Re: [PATCH v2 4/4] copy_signal cleanup: clean tty_audit_fork()

[Oleg Nesterov <oleg@redhat.com>]

On 12/06, Miloslav Trmac wrote:
>
> ----- "Oleg Nesterov" <oleg@redhat.com> wrote:
> > On 12/05, Miloslav Trmac wrote:
> > > > Off-topic question to this who understands this code.
> > > >
> > > > But afaics we can also remove ->siglock from this helper and make
> > > > it really trivial for being inline. ->siglock buys nothing, we just
> > > > read a boolean. In fact, after the quick grep I do not understand
> > > > how ->siglock is connected to ->audit_tty. OK, it protects
> > > > tty_audit_buf,
> > > > but why we always take ->siglock to access ->audit_tty ?
> > > AFAIK there is no explicit documentation of the atomicity semantics
> > > expected by the Linux kernel (both from the hardware and from the compiler),
> > > so every access to the boolean is protected by a lock, to be on the safe side.
> >
> > Not sure I understand, but the kernel relies on fact it is always safe
> > to load/store a word.
> And is "word" an "unsigned", "unsigned long" or "intptr_t"?  Must it be
> suitably aligned, and if so, what is "suitably"?

Sure, it must be aligned.

> Where is this documented?

Perhaps nowhere, I do not know. If this is not documented, probably
it would be nice to add a note.

> > What atomicity semantics do you mean and how ->siglock can help?
> At the very least, "any access will read the last value stored and not result
> in undefined behavior, even if other threads attempt to access the value".
> In user-space, per POSIX, the only way to guarantee this is using explicit
> synchronization primitives.

We have numerous examples in kernel code which rely on this fact.

If we are talking about copy_process() pathes, please look at, say,
sched_fork(). Say, we read current->normal_prio lockless, while another
thread could change ->normal_prio in parallel.

Oleg.