On 05/03, Peter Zijlstra wrote:
>
> On Thu, May 02, 2019 at 12:09:32PM +0200, Oleg Nesterov wrote:
>
> > > +static void readers_block(struct percpu_rw_semaphore *sem)
> > > +{
> > > + wait_event_cmd(sem->writer, !sem->readers_block,
> > > + __up_read(&sem->rw_sem), __down_read(&sem->rw_sem));
> > > +}
> > > +
> > > +static void block_readers(struct percpu_rw_semaphore *sem)
> > > +{
> > > + wait_event_exclusive_cmd(sem->writer, !sem->readers_block,
> > > + __up_write(&sem->rw_sem),
> > > + __down_write(&sem->rw_sem));
> > > + /*
> > > + * Notify new readers to block; up until now, and thus throughout the
> > > + * longish rcu_sync_enter() above, new readers could still come in.
> > > + */
> > > + WRITE_ONCE(sem->readers_block, 1);
> > > +}
> >
> > So iiuc, despite it name block_readers() also serializes the writers, ->rw_sem
> > can be dropped by down_write_non_owner() so the new writer can take this lock.
>
> I don't think block_readers() is sufficient to serialize writers;
> suppose two concurrent callers when !->readers_block. Without ->rwsem
> that case would not serialize.
Of course. I meant that the next writer can enter block_readers() if
up_non_owner() drops ->rw_sem, but it will block in wait_event(!readers_block).
(And if we change this code to use wait_event(xchg(readers_block) == 0) we
can remove rw_sem altogether).
The main problem is that this is sub-optimal. We can have a lot of readers
sleeping in __down_read() when percpu_down_write() succeeds, then after
percpu_down_write_non_owner() does up_write() they all will be woken just
to hang in readers_block(). Plus the new readers will need to pass the
lock-check-unlock-schedule path.
Peter, just in case... I see another patch from you but I need to run away
till Monday.
Oleg.
